"Categorization is the mental operation by which the brain classifies objects and events. This operation is the basis for the construction of our knowledge of the world. It is the most basic phenomenon of cognition, and consequently the most fundamental problem of cognitive science."

from the intro paper in

Cohen, H., & Lefebvre, C. (Eds.). (2005). Handbook of categorization in cognitive science. Elsevier.

I quote the above because I have yet to find a more concise, poignant, and apt expression of categorization’s fundamental importance to the entirety of the cognitive sciences - including cognitive and computational neuroscience. Unfortunately, computational neuroscience deals with categorization largely through methods like neural network models. While these are fantastically successful in many ways, they are quite fundamentally incapable (at the moment) of anything resembling an explanation for conceptual categories (rather than discrimination and other ‘learning” mechanisms). B.A. Olshausen’s contribution to the volume “20 Years of Computational Neuroscience” provides another useful, concise statement: “At the end of the day we are faced with this simple truth: No one has yet spelled out a detailed model of V1 that incorporates its true biophysical complexity and exploits this complexity to process visual information in a meaningful or useful way . The problem is not just that we lack the proper data, but that we don’t even have the right conceptual framework for thinking about what is happening.” (the entire sentence after “this simple truth” is in italics).

That volume was published in 2013, and alas qualitative advances have not occurred over the last ~year. Certainly, there are hypothetical mechanisms numerous studies have implicated, such as synchronization, but here too often there are questions that must be answered before we can even test such hypotheses: “the mechanism for the emergence of correlation, synchronization, or even nearly zero-lag synchronization (ZLS) among two or more cortical areas which do not share the same input is one of the main enigmas in neuroscience. It has been argued that nonlocal synchronization is a marker of binding activities in different cortical areas into one perceptual entity”. The second sentence is a hypothesized mechanism for perceptual categorization, but as the first points out it the mechanism itself is a puzzle.

On the flip side, cognitive neuroscience(s) do address conceptual/perceptual classification and categorization. However, here we find the opposite issue as that faced in computational neuroscience: neural processes are largely ignored and the focus is on comparatively massive brain regions. Without the mathematical rigor provided by a computational approach, naturally the dependency upon numerous assumptions in any given experimental design, let alone interpretations of results, presents additional problems. Perhaps the best illustration is the “embodied model of cognition vs. the largely incompatible “classical” (symbol-processing & modality-specific symbol representation) model of cognition. The differences between the two models mean that researchers disagree over the ways in which categorization should even be approached.

Additionally, functional neuroimaging technologies in general and functional MRI in particular are the central tools used in the neurosciences for research on categorization. Yet disagreements exist at every level of their use (see e.g., the volume Foundational Issues in Human Brain Mapping and the papers therein).

Finally, a great deal of neuroscience research (excluding computational) involves not only signal processing of high dimensional NMR data (simply put- quantum mechanical technology that orients otherwise “randomly” distributed alignment of spin in hydrogen atoms) used as a proxy for hemodynamic activity itself used as a proxy for cognitive processing all of which is almost universally ignored in all fMRI studies. But for one issue this wouldn’t matter at all. These data require sophisticated statistical methods to render into something that researchers can use to relate the data garnered from e.g., participant responses, choices, reaction times, etc. However, education in the behavioral & social sciences has not caught up with the current need for advanced mathematical modelling, data analyses, and statistical methods. As a result, researchers rely on software packages capable of easily doing all of the above without understanding the underlying logic. “Garbage in, garbage out”.

That said, we have made substantial strides in our understanding of categorization. A great deal of it, however, comes not from neuroscience research but e.g., cognitive psychology and linguistics. There is certainly a great deal of (I believe) sound research within the neurosciences but once we move beyond the categorization that single cells and sea slugs are capable of, this research contains a large amount of problematic studies building upon other problematic studies. I do not think we are yet at the point where we can say much with certainty regarding the ways in the brain categorizes perceptual or conceptual information (not the two are distinct; they overlap at many levels regardless of whether one is a proponent of embodied cognition).

I have to attend a birthday party, so I apologize for such a negative response rather than answers, but I will make up for that ASAP.

Dear Joachim, many thanks for your succinct answer and for all of the references. Is this a major part of your work? I cold not see from your page where you work?

Joachim, do you have sources for the following claims:

"Categorization is something every feed forward neural network does,

especially when there are only few neurons so that training patterns

(input / class pairs) can't be learned rote. Then the neural net

interpolates, creates categories. On a higher, symbolic

level there are state machines, parsers, and pattern matchers"

"Categorization is the mental operation by which the brain classifies objects and events. This operation is the basis for the construction of our knowledge of the world. It is the most basic phenomenon of cognition, and consequently the most fundamental problem of cognitive science."

from the intro paper in

Cohen, H., & Lefebvre, C. (Eds.). (2005). Handbook of categorization in cognitive science. Elsevier.

I quote the above because I have yet to find a more concise, poignant, and apt expression of categorization’s fundamental importance to the entirety of the cognitive sciences - including cognitive and computational neuroscience. Unfortunately, computational neuroscience deals with categorization largely through methods like neural network models. While these are fantastically successful in many ways, they are quite fundamentally incapable (at the moment) of anything resembling an explanation for conceptual categories (rather than discrimination and other ‘learning” mechanisms). B.A. Olshausen’s contribution to the volume “20 Years of Computational Neuroscience” provides another useful, concise statement: “At the end of the day we are faced with this simple truth: No one has yet spelled out a detailed model of V1 that incorporates its true biophysical complexity and exploits this complexity to process visual information in a meaningful or useful way . The problem is not just that we lack the proper data, but that we don’t even have the right conceptual framework for thinking about what is happening.” (the entire sentence after “this simple truth” is in italics).

That volume was published in 2013, and alas qualitative advances have not occurred over the last ~year. Certainly, there are hypothetical mechanisms numerous studies have implicated, such as synchronization, but here too often there are questions that must be answered before we can even test such hypotheses: “the mechanism for the emergence of correlation, synchronization, or even nearly zero-lag synchronization (ZLS) among two or more cortical areas which do not share the same input is one of the main enigmas in neuroscience. It has been argued that nonlocal synchronization is a marker of binding activities in different cortical areas into one perceptual entity”. The second sentence is a hypothesized mechanism for perceptual categorization, but as the first points out it the mechanism itself is a puzzle.

On the flip side, cognitive neuroscience(s) do address conceptual/perceptual classification and categorization. However, here we find the opposite issue as that faced in computational neuroscience: neural processes are largely ignored and the focus is on comparatively massive brain regions. Without the mathematical rigor provided by a computational approach, naturally the dependency upon numerous assumptions in any given experimental design, let alone interpretations of results, presents additional problems. Perhaps the best illustration is the “embodied model of cognition vs. the largely incompatible “classical” (symbol-processing & modality-specific symbol representation) model of cognition. The differences between the two models mean that researchers disagree over the ways in which categorization should even be approached.

Additionally, functional neuroimaging technologies in general and functional MRI in particular are the central tools used in the neurosciences for research on categorization. Yet disagreements exist at every level of their use (see e.g., the volume Foundational Issues in Human Brain Mapping and the papers therein).

Finally, a great deal of neuroscience research (excluding computational) involves not only signal processing of high dimensional NMR data (simply put- quantum mechanical technology that orients otherwise “randomly” distributed alignment of spin in hydrogen atoms) used as a proxy for hemodynamic activity itself used as a proxy for cognitive processing all of which is almost universally ignored in all fMRI studies. But for one issue this wouldn’t matter at all. These data require sophisticated statistical methods to render into something that researchers can use to relate the data garnered from e.g., participant responses, choices, reaction times, etc. However, education in the behavioral & social sciences has not caught up with the current need for advanced mathematical modelling, data analyses, and statistical methods. As a result, researchers rely on software packages capable of easily doing all of the above without understanding the underlying logic. “Garbage in, garbage out”.

That said, we have made substantial strides in our understanding of categorization. A great deal of it, however, comes not from neuroscience research but e.g., cognitive psychology and linguistics. There is certainly a great deal of (I believe) sound research within the neurosciences but once we move beyond the categorization that single cells and sea slugs are capable of, this research contains a large amount of problematic studies building upon other problematic studies. I do not think we are yet at the point where we can say much with certainty regarding the ways in the brain categorizes perceptual or conceptual information (not the two are distinct; they overlap at many levels regardless of whether one is a proponent of embodied cognition).

I have to attend a birthday party, so I apologize for such a negative response rather than answers, but I will make up for that ASAP.

As promised, some actual (and hopefully useful) sources:

First, some relevant talks/presentations you may find useful:

"Computational Analysis Methods and Issues in Human Cognitive Neuroscience" given by Dr. Bradley Voytek:

http://youtu.be/nZRXrOq1T08

"The Cognitive and Computational Neuroscience of Categorization, Novelty-Detection and the Neural Representation of Similarity" given by Dr. Mark A Gluck

http://youtu.be/2Ei6wFJ9kCc

"Untangling object recognition: Which neuronal population codes can explain human object recognition performance?" by Dr. Jim DiCarlo

http://youtu.be/d7eUonmK82E

Second, it is important to distinguish between classification of the type made by e.g., feedforward neural networks algorithms, support vector machines, statistical learning algorithms (like Bayesian networks), kernel methods, etc. These are all computational methods with applications far beyond neuroscience models whether they are biologically inspired or not, but the "type" of classification they are used for is of the same sort used by single-celled organisms, plants, & sea slugs (Aplysia californica) are capable of. Call it syntactical processing, procedural learning, habituation, or whatever you wish, it is the system's purely reactionary response to stimuli by adaption. The system's integration of information (reaction to stimuli or input) is not conceptual nor is the representation of this information categorical. Rather, the classification mechanisms are simply the adaptive/reactive processes through which the system is able to attain a particular state which is itself the representation of the "class". There is no generalized concept abstracted away from specific input. Concepts are classes/categories which, even when they are physical entities (trees, cars, dogs, etc.) are abstracted away from any specific instantiation. A central goal of neuroscience (and the cognitive sciences in general) is to understand how species from mice to humans are able categorize perceptual information according to conceptual (dynamic) interrelated networks.

It is well known that categorization and conceptualization involve "fuzzy" boundaries. Many neuroimaging studies (among other methods) seek to determine what properties determine whether humans will categorize something as a dining utensil vs. a kitchen tool or a car vs. a van or far more broadly what it is humans and other animals use to determine similarities and to what extent certain similarities (e.g., color or shape) in some context are les important than others (e.g., functionality, some feature of a conceptual exemplar/category prototype, etc.).

At the moment, I would still argue that most of what we know and draw upon to expand our knowledge doesn't come from neuroscience per se, but from cognitive psychology, cognitive linguistics, developmental psychology, etc. I was given From Molecule to Metaphor: A Neural Theory of Language by J. A. Feldmen for Christmas. While it was intended for the general reader, I still found it an interesting read in part because before we reach the first chapter two highly contentious claims have already been made. First, the title hearkens back to Lakoff & Johnson's Metaphors we live By and Lakoff's Women, Fire, & Dangerous things as well as a core aspect of cognitive linguistics: the idea that metaphor is central to cognition. The other was that Chomsky was basically alone in not believing this to be true. Most importantly, though, this book drew upon linguistic theories developed in the 80s and 90s at least as much and probably more than it did neuroscience.

Slightly OT but also in the social sciences types or categories have been long recognized as fundamental, for example by Schütz who stated that all common sense but also science is based on types (categories): Schutz, A. (1953) Common-sense and scientific interpretation of human action. Philosophy and Phenomenological Research 14, 1-38.

Thanks Johannes, it is interesting to see seminal work in this area. What does OT mean?

Thanks Andrew, I will think on what you have said.

it means off topic as you were asking for neurological basis and I was bringing in social sciences ;)

Thanks Johannes but what you said was not at all OT!

I too am off topic. My question concerns the prevalence of repeated myths and archetypal figures across cultures. I have a hunch, that categorization begins with the Platonic idea of the perfect form from which all others are compared and collected. My theory is that archetypes and myths are emanations from the primary brain that set up the conditions for their existence. As a non-scientist, how bizarre is my thinking or have I stumbled on to something?

I don't know if it was deliberate, but I have to say I find quite intriguing your phrase "Platonic idea of the perfect form" (clever, funny, sophisticated, and/or some other set of adjectives that I can't recall at the moment). In Greek, the word for Plato's "Forms" is usually εἰδέα (plural of εἶδος), but another lexeme he used was ἰδέα. Standard transliteration would have us render ἰδέα as "idea". Thus, the "Platonic ἰδέα of forms is, in a way, the Plato's form of ideas. (I'm too drunk and tired to recall the proper grammatical way to combine to genitive constructions out of the nominative ἰδέα, and Πλάτωνος & εἰδῶν).

I wrote a paper entitled something like "Chomsky’s Failure: Malignant Platonism in Cognitive Science". It was a while ago and in no small measure intended to be sensationalist. A. N. Whitehead once remarked that the whole of Western philosophy is but a "series of footnotes to Plato". In this case, I think it is an unfortunate truth. I have to admit that despite my arrogant disdain for all things Jungian, I never thought to reach into Freud's disciples when considering the historical context of language, concepts, and categorization.

Plato's forms dominated our conceptualization of conceptualization for centuries. Lakoff, in his 1987 Women, Fire, and Dangerous Things: What Categories Reveal About the Mind, put forth an as yet unrivaled critique of the Platonic approach. The title of the book (at least the main part) was either hated or loved by feminists for precisely the same reason: it suggests to us that Lakoff was grouping the words "women", "fire', and "dangerous things" into a category. In a sense, he was: the category is a word from an Australian language, but Lakoff's point was to show how our conceptions of categorization being a those elements which are instantiations of a perfect ideal are too ideal, too neat. We do not categorize so precisely. A pencil and a string bean have much the same shape, yet most would probably group "pencil" in which marker, chalk, even chalkboard rather than in any grouping that included "string bean". Similarly, "string bean" would be categorized alongside vegetables of all shapes and colors.

That said, "exemplars' do bear more than a passing similarity to archetypes (and were suggested as a categorization mechanism precisely because prototypes were too precise to account for data).

This I will have to think on when it is not 4 in the morning. Thank you!

Marvelous!

The work exemplars is perhaps the line I am following. And, I postulate that archetypes can be considered exemplars therefore of impulses to action whether considered good or bad.

Typo "work" is wrong; I meant "word". The word exemplars is perhaps the line I am following. And, I postulate that archetypes can be considered exemplars therefore of impulses to action whether good or bad, the essence of lust, rage, etc. These essences are given faces called gods.

what happens to the regions of proteins which do not participate in fibrilation of proteins:

In many proteins such as lysozyme and alpha- synuclein ,only a limited part of protein participate in cross-beta sheet and fibril formation. In alpha synuclein from 140 residues only about 60 residues are present in fibrils. We know that profibrils have lateral intraction and lateral non firilar regions can interfere with this lateral intraction. We saw that in alpha-synucleion, in all stages of fibrilation labeled-N-terminal part of protein remained. how?

The “category formation” question may beg the very puzzle it is designed to solve, namely how do we recognize objects and events, use our past experience with them to respond, to think, to imagine and so forth. The term “category” implies a discrete property that is often incompatible with perception, imagination, problem solving, etc. I’m reminded of a world famous physicist and linguist who upon introducing a famous Japanese physicist to his graduate seminar, asked him, “But why do you mispronounce your name?” The Japanese pronunciation didn’t fit accepted Japanese linguistic categories!

So I suggest that we continue to model cognitive processes such as recognition and problem solving and to examine the category-like representations that are necessary to accomplish these processes. A good place to start is to build a visual word recognition model with the connectionist Emergent model of O’Reilly. http://grey.colorado.edu/emergent/index.php/Main_Page

As you will see at least one set of “hidden” “categories” are needed to recognize letter (categories) from visual letter features (categories), and an addition “hidden” categories to recognize word (categories) from letters. I think a little experience with such a connectionist model may help to reframe the “category formation” question. For starts, it will help to understand how higher level categories such as words and constructed from several layers of more elementary categories, many of which, unlike letters, we are unaware.

Mr Antrobus (speaking of names) Tennessee Williams appropriated yours for "The Skin of Our Teeth." But I suspect you knew that. As the holder of an MFA, and not even a PhD, I tried to look at the Emergent model link and got no where. My old brain couldn't parse it out. I suspect that the "hidden categories" to which you refer at the elementary levels are probably based in the primal brain areas. Freeing actors to respond to their impulses, means allowing them to access such urges. My theory is that language use is an attempt to negotiate between wild and inappropriate needs and social acceptance. Speaking rather than lunging for something or grabbing someone's hair, or running and hiding, is a means of protection even when the text may seem to be a personal revelation. But, I suspect I missed the train and that everyone has already figured that out or rejected it. This is one of the reasons that I have joined RG, I really want to know how my experiential learning lines up with current scientific thought.

I think it's important to put "connectionism" into some historical context. This was the first (and still cited, still used) connectionist model:

McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. The bulletin of mathematical biophysics, 5(4), 115-133.

It provided the mathematical foundation for connectionism, for "neural networks" before the neurophysiological foundation. In fact, the neurophysiological contribution to connectionism is mentioned often, but not so much in the context of connectionism. Nor was it the next huge contribution. The next was

Hebb, D. O. (1949). The organization of behavior: A neuropsychological approach. New York: Wiley.

The above book is the reason for the term Hebbian learning. It was, once again, a contribution to the mathematical formulation of associative, distributed learning, and it came before the first biologically-based (and neurophysiological) model of learning:

Hodgkin, A. L., & Huxley, A. F. (1952). A quantitative description of membrane current and its application to conduction and excitation in nerve. The Journal of physiology, 117(4), 500.

All of this before Chomsky had written Syntactic Structures, before Fodor turned 20, while Shannon and Skinner's behaviorism (and Skinner himself) were still powerhouses in the approach to the brain, and before cognitive science really existed.

Then came Chomsky, Marr, etc., and the symbolic-algorithmic (or computational) model. The growth of computers, the increasingly sophisticated emerging computer sciences, etc., and strong AI (human like consciousness) was just around the corner. The brain was just another processor, manipulating rules and all we had to do was to figure out the algorithms. After about 30 years (early 80s), the early work cited above was pulled to the forefront with works like e.g.,

Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences, 79(8), 2554-2558.

And it was really exciting. And strong AI was just around the corner- again. We main computers act like brains with connectionism's neural networks and similar adaption-based algorithms (ANNs, swarm intelligence, evolutionary algorithms, gene expression, fuzzy logic, etc.). We used biology as our inspiration, threw out the rigidly formal procedural approach of traditional algorithms, replaced it with algorithms that determined how programs ADAPTED to input, and in general made things messier (like living systems) to be more powerful and more flexible.

The result? Sea slug learning. I use that term partly because I am so tired of the popular science heralding of the singularity/strong AI/etc., but also for another: a central contribution (Nobel prize winning, in fact) to adaption-based learning was Erik Kandel's work with sea slugs (Aplysia californica). That's what connectionism provided. A way to "learn" by adaptive algorithms such that input that would change the system's output rather than determine it. It works fantastically for all sorts problems: recommending ads Amazon, the way Google or your smart phone guesses what you are trying to type in before you do, the way Netflix recommendations work, etc. It provided nothing to our understanding of conceptualization and the mechanisms employed by the most sophisticated, custom-built AI machines with state-of-the-art computational intelligence algorithms are still using sea slug learning.

The assertion that "categories' implies something "discrete" is mostly true. However, it isn't, in general, true in the cognitive sciences. Fuzzy sets date back to Zadeh's work in the 60s. "Cognitive linguistics" was founded by work in the 80s and 90s (actually, it really started in the late 60s and early 70s), and is an umbrella framework for a set of related approaches to language and cognition that include "fuzzy", over-lapping categories, polysemy, the vital importance of embodiment and metaphor for even abstract cognitive processing, etc. (and rejects Fodor's massive modularity, Chomsky's universal grammar, Pinker's domain-specificity, etc.). Work on conceptual categories that are neither discrete nor well-defined have existed in linguistics and psychology (for psychology, see Eleanor Rosch's work on prototype theory as well as a good deal of related work in Gestalt psychology).

The neural network simulator "emergent", like NEURON, like its predecessors, and like other similar environments/simulators runs on a finite state machine that is equivalent to any Turing machine. Emergence, however, ill-defined it is (like complexity), is characterized by the ancient concept of "the whole is greater than the sum of its parts". A computer is exactly the sum of its part and is designed precisely to be that way. That's why it does what it does so well: compute. Everything is very efficient, very neat, very discrete. Whatever we do to make it more like information processing in living systems we do by using the essential architecture of computers against themselves.

Criticisms (well-founded or not) against the idea that living systems are computable (per Rosen and systems biology) or that the mind is a computer have been around since Turing himself. There are many to choose from (expressed in various ways and by no means all wholly critical), but among the many formulations I typically turn to when I wish to represent the opposition to any kind of "computable mind" is from Torey's Crucible of Consciousness:

"No mind inquiry is complete without examining the position of analytic philosophy and computer science. The symbolic operations with which they are concerned are marginally cognate with one particular aspect of thinking, and the overlap has been taken to imply that the mind resembles these operations or that these operations are analogues of the mind. This section demonstrates that there is no similarity and that no formal system, be it logic, mathematics, or analytic philosophy, is mind-like or a source of insight into the nature and structure of the mind. The task is to show that it was the brain that generated formalism, such as grammar and logic, and that no formal system is able to generate anything even remotely mind-like. The asymmetry between the brain and the computer is complete, all comparisons are flawed and the idea of computer-generated consciousness is nonsense."

The middle ground between computational neuroscience (a formal treatment of neuronal activity that is as ill-suited to modelling informal concepts as the descriptor "formal" implies) and cognitive psychology or cognitive neuroscience is still lacking. Whatever is involved in conceptual processing, I don't think we'll get very far with computer programs that rest upon a few centuries of creating formal languages to rid mathematics of the ambiguities so fundamental, integral, and ubiquitous to language and thought.

The neurological basis of category formation is explicitly proposed in terms of the properties of a particular kind of neuronal mechanism -- the synaptic matrix. For details see "Learning, Imagery, Tokens, and Types: The Synaptic Matrix" on my RG page.

The problems raised in this thread are addressed in some detail in *The Cognitive Brain* (MIT Press 1991). For relevant simulation tests of the brain mechanisms proposed see "Self-Directed Learning in a Complex Environment" on my RG page.

The neurological basis of category formation has been proposed explicitly for the past ~50 years in numerous ways. It's been proposed as irrelevant for just as long (see e.g., Boden's 2-volume "Mind as Machine" for a historical perspective and the four volume "Learning and Memory: A Comprehensive Reference" by Menzel & Byrne and the volumes of "Advances in Brain Research" for examples). Alwyn C. Scott is as much a hero to me as any contributor to complexity sciences (or any science, for that matter), but even his posthumously published "The Nonlinear Universe" refers us to McCulloch & Pitts, Hebb, and Aristotle. The problems in this thread have been addressed in detail for the past ~2,500 years. We now have proposals that the answers to human conceptual categorization are to be found in quantum physics, are to be found using the mathematical apparatus of quantum physics, are algorithmic and the neurophysiology is largely (or entirely) irrelevant, are to be understood as feature specific (vs. category-specific or more recently process specific), are to best approached through a systems perspective, are best approached through a dynamical systems phase space/geometrical perspective, and so forth. What we don't have is any model that has allowed us to demonstrate how conceptual processing works. I would suggest that those who are perhaps less familiar with the concept of categorization as it has existed in the cognitive sciences since their inception not limit themselves to one of hundreds of explicit proposals, but perhaps seek more diverse sources (such as the Handbook of Categorization in Cognitive Science). It's hard to evaluate the veracity of any single model without a background in what kinds of models exist and why as well as the relations between them, let alone appreciate the value of one such as the synaptic matrix.

Andrew, among the neuronal models that you know about, which have competence for parsing, learning, and searching for objects in a complex environment? I would also be interested in knowing your selection of a neuronal model that can operate as effectively as the one described in "Sparse Coding of Faces ... " on my RG page.

That depends upon one's view of what constitutes "neuronal model" (i.e., whether one holds that a model needs to be biologically-based and, if so, perhaps also one's beliefs about the neurophysiological processes would make a model accurately "biologically-based" vs. based on a model of neuronal function that one disagrees with). Basically all neural networks around are effective at "parsing, learning, and searching for objects in a complex environment". For example:

Unsupervised Learning of Categorical Segments in Image Collections

(http://www.vision.caltech.edu/publications/POCV08.pdf)

or better yet:

Narwaria, M., & Lin, W. (2010). Objective image quality assessment based on support vector regression. Neural Networks, IEEE Transactions on, 21(3), 515-519.

The latter paper so vastly outstrips any kind of visual classification algorithms around when that volume (121) of Advances in Psychology came out that it categorizes not the image but its quality.

However, no effort is made (or attention paid) to biological plausibility as, like most biologically-inspired computational intelligence/soft computing approaches, neural networks are mostly used outside of computational neuroscience.

If we equate neural networks with neuronal models in this way, than we are left with very little cause to exclude e.g.,

Robust face landmark estimation under occlusion

(http://vision.caltech.edu/~xpburgos/papers/ICCV13%20Burgos-Artizzu.pdf)

The above, however, involves a different kind of machine learning. For a plethora of such work (some actually concerned with biology) Caltech's site is probably best:

http://www.vision.caltech.edu/publications/publications.html

For one thing, regardless of methods/models in computer science vs. neuroscience, Caltech-256 (and before that, Caltech-101) is among the most common benchmark (see Visual Recognition Circa 2008 by P. Perona in "Object Categorization- Computer and Human Vision Perspectives"). Yet despite the fact that "a tremendous amount of effort is being expended on exploring algorithms that may support general object recognition...one can hold up an actual “product” and give performance numbers, which feels like progress. Indeed, to the newly initiated, it often sounds as if such algorithms have solved the problem of object recognition, because successes are highlighted more than failures...We do not wish to impugn the very creative work in the community...However, it is not clear that such a piecemeal approach will lead us to a general theory of object recognition, and the short-term goal of succeeding on limited problems may distract us from the crux computational issues" (from the paper immediately prior to Perona's, "A Strategy for Understanding How the Brain Accomplishes Object Recognition" by J. J. DiCarlo).

For another benchmark that also critiques current approaches:

Establishing Good Benchmarks and Baselines for Face Recognition

(http://hal.inria.fr/docs/00/32/67/32/PDF/pinto-dicarlo-cox-eccv-2008-lfw_final.pdf)

Modern work on visual system models not only all fit your description but they are beginning to include hardware that models some theoretical hierarchy of the visual system. See e.g.,:

Bharath & Petrou (Eds.). (2008). Next Generation Artificial Vision Systems-Reverse Engineering the Human Visual System

Perelman, Y., & Ginosar, R. (2008). The Neuroprocessor: An Integrated Interface to Biological Neural Networks. Springer.

Rasche, C. (2005). The making of a neuromorphic visual system. Springer.

Human-Centric Machine Vision (available online for free here: http://www.intechopen.com/books/human-centric-machine-vision), although I include this one just in case your access to the others isn't ready enough to warrant the effort required obtain them.

I have to admit that once I turned from languages and psychology to computational sciences, mathematics, and complexity (centered, of course, around the brain and computational & cognitive neuroscience), I spent a long time obsessed with mathematical models, computational intelligence paradigms, machine learning, etc., without thought to whether or not it was useful for understanding anything other than what applied problems could be solved or simply the elegance of the mathematics. I would like to think I've moved beyond that somewhat. As I said in an earlier response, I think it best to distinguish between classification (or discrimination) of the type our best computational methods make and conceptual categorization. Neuronal models that are capable of "learning" the way that single cells or plants are "learn" will not, I think, end up explaining how conceptual categorization works. To repeat a quote from that response:

“At the end of the day we are faced with this simple truth: No one has yet spelled out a detailed model of V1 that incorporates its true biophysical complexity and exploits this complexity to process visual information in a meaningful or useful way. The problem is not just that we lack the proper data, but that we don’t even have the right conceptual framework for thinking about what is happening.”

That quote comes from B. A. Olshausen paper "20 Years of Learning About Vision: Questions Answered, Questions Unanswered, and Questions Not Yet Asked" (his contribution to the volume cited earlier).True, his historical perspective pales in comparison to our own. Although I do own your most well-known book as well as that Advances in Psychology volume, it is mainly through these that I am aware of your work in the 70s and, if memory serves, a monograph (?) from the 1950s (?), and so I realize that while Olshausen's description of progress before the "20 years" he can speak to (back to the 60s) is like mine: informed by secondary written and human sources.

However, as the quote comes from the section of his paper that deals with what is still not known, the lack of historical perspective does not, I think, play much of a role.

I do not think that most neuronal networks, no matter what their ability to navigate some environment, classify input, learn linguistic rules, etc., are telling us much. I own perhaps 1 or 2 dozen books/volumes that could be said to have a neuroscience orientation and include in their title "neural network". I have many times that number that are concerned with machine learning or soft computing. And even those in series like Springer Series in Cognitive and Neural Systems include large sections on computer science, engineering, and similar atheoretic approaches to neuronal models.

Interestingly, one such volume (The Relevance of the Time Domain to Neural Network Models) includes a paper you might find interesting "Temporal Coding Is Not Only About Cooperation—It Is Also About Competition". Like your model, it involves competition. However, it does so by incorporating research in neuroscience since your paper was published on neural encoding and synchronization. Although the modelling discussed isn't backed by any report of some training set results, the equations are thoroughly inline with neurophysiology research and generally known properties of dynamical systems and synchronization (on the importance of the latter, see e.g., Pikovsky, A., Rosenblum, M., & Kurths, J. (2001). Synchronization: a universal concept in nonlinear sciences).

From a computational perspective, my interests lie more with mathematical modelling that have more in common with systems biology, biophysics, and quantum mechanics (no, I don't by into quantum consciousness). In part this is because a paper like "Nonlocal Mechanism for Synchronization of Time Delay Networks" not only investigates "the hypothesis that synchronization may hold key information about higher and complex functionalities of the network" but relies at least as much on physics as it does on neurobiology (rather than neurobiology and computer science). In part it is because I find idea like "massive modularity", hierarchies, and the comparison between the kind of discrimination/classification neural networks and similar models are capable of to by non-analogous to anything other than procedural memory. Most of all, though, it is because of the relatively extreme lack of progress in the life sciences (including neurobiology) relative to advances in understanding complex systems in general. While Rosen and his followers are arguing that living systems must have non-computable models and their opposition proposing everything from hypercomputation to exotic logics in defense of such computable models, many in the neurosciences are relying on mathematically outdated methods to assumptions based on an understanding of the brain heavily dependent (if only implicitly through historical influences) upon computer science and obsolete physics. If I'm wrong, great. I'd love to be. I'd love to be able to write programs that can categorize the way my dog can instead of the way a plant can. Currently, nobody can. I spent several years trying to show that quantum consciousness is not just without any empirical support but cannot be a model for consciousness, which I still maintain. However, the deterministic, reductionism that so defines the brain sciences is incompatible with much of modern physics and, whether the approach is a good one or not, there is no debate within the brain sciences about such issues compared to the debates within modern physics and (some of) the life sciences. Also, what discussion does exist is mostly motivated by outsiders.

But enough of my incoherent ramblings. I'll await your (justified) response asking for me to actually answer the questions you asked that I've drifted so far from I'm sure I have not.

So, Paul, it appears that the answer to your question embraces all of cognitive neuroscience - which is still in its infancy. It might help if you gave specific examples of the processes of interest to you. Often it is not helpful to try to address them with concepts that we have carried from the early Greeks. Category formation may not provide you with the best answer.

Joachim. Thanks for the website information. Just one comment. Although nearly all theorists treat recognition as a bottom-up or feed forward process - and it is if our models use times units larger than 1 sec., and don’t examine what happens within that 500 ms - 1 sec interval. But in fact, all recognition is a complex sequence of interacting top-down-bottom up and lateral processes. Some parts of the recognition system have 10 times as many descending as ascending neurons!! What do they do? All recognition occurs in complex contexts, the “categories” of which bias every sec. of real world recognition by the top-down, and lateral, legs of the recognition process. Much of the bias is in changes in synaptic strength among the participating neurons. An example - in the context of this discussion, the V1 cortex has to process only the first 2 letters of your opening sentence to “recognize” the word “Categorization.”

John: "An example - in the context of this discussion, the V1 cortex has to process only the first 2 letters of your opening sentence to “recognize” the word “Categorization.”

Are you claiming that the neuronal mechanisms in V1 actually recognize words? Or do you mean that a word is often recognized on the basis of presentation of the first two letters of a word?

The last sentence only. See Cloze technique research - . In order to maximize cortical recognition efficiency, if the word can be recognized (in the left latter cortex word recognition region) with only the first letter or so, then all the costly time consuming lower level processes can be eliminated - from V1 up. This frees the cortex to process information of greater utility. Without this array of top-down context recognition biases, every sec. of waking perception would be like waking up in a foreign hotel room!

John: "Without this array of top-down context recognition biases, every sec. of waking perception would be like waking up in a foreign hotel room!"

I assume that the "array of top-down context recognition biases" are the result of parallel pre-conscious semantic associations. They certainly affect the meaning of words but they would not change our *perception* of the current environment. In fact, our current perception of the world influences whether a given word in current context makes sense.

Thank you all for your comments and suggestions. I will be reading works that you suggested and will then get back to you? Thanks

Arnold. I assume that prior recognition of the word "categorization" in the V1-to-left temporal cortex visual lexical word recognition region goes on to activate an array of interacting semantic association circuits in the prefrontal cortex which in turn recycle for 5-15 minutes, during which time they also send actoivation back to the top of the visual lexical word in the temporal cortex. Not too much or it would mess up the recognition of subsequent words in the text, but enough to maintain a modest resting level for that word. That resting activation inhibits the activation of competing words beginning with "ca.." such as "cattle" and "catastrophy," which are also activated by the "ca.." in the bottom-up visual word pathway. Consequently, "ca.." is sufficient to fully activate the correct partially viewed word. ... makes reading the newspaper a much faster job! I call this proception. See the work of Moise Bar on the proactive brain.

To return somewhat to neurophysiology, (neuronal) network topology, and synchrony, it seems somewhat relevant to situate our understanding of the brain within our understanding of complex systems consisting of many interrelated and complex networks. For those who follow such things, the concept of "network motifs". Simplistically, motifs can be described statistically, characterized by recurrent patterns, understood via the bounded behavior of the system's evolution within its phase space, or perhaps most aptly (given our context) as modularity within large, complex systems. In other words, current research suggests that perhaps some complex networks (such as those formed by social networking sites or even the internet itself) are characterized by "smaller" (localized) repetitive or near-repetitive behavior of their components that allow us to semi-reduce such networks to the approximate dynamics of constituent subsystem component networks. Roughly speaking, this is akin to approaching the brain in terms of massive modularity, domain specificity, and in general a form of compartmentalization that (despite obvious interconnectedness among "distinct" neuronal networks) can be treated singularly without significant loss of generality.

Bar's work (and similar work) seem to both support such a model and reject it, but I think this is mainly because the "proactive brain" is described qualitatively and neural processes merely suggested as possible mechanisms rather than proposed as explanatory models. A possible issue is equating "learning" (of the Hebbian, "fire-together-wire-together", habituation, etc.-type) with localized associations between networks. It can be useful in many ways for to treat visual networks as at least approximately independent of e.g., cortical networks and as characterized by repeated, "learned" patterned responses to stimuli that, to the extent they are modulated by inter-network (global) reactions they are minimally so.

Thus Bar's description of mediating systems (MTL, PHC, etc.), is in one sense a synchronic approach ("The neural mechanism that generates predictions is largely unknown, but it is thought to be mediated, or at least balanced, by neural oscillations and synchrony"), but of a particular type. Once more, I think it useful to incorporate work in complex/nonlinear systems and physics, and in fact we find useful progress here:

Kopelowitz, E., Abeles, M., Cohen, D., & Kanter, I. (2012). Sensitivity of global network dynamics to local parameters versus motif structure in a cortexlike neuronal model. Physical Review E, 85(5), 051902.

Like sources I mentioned earlier, the authors examine nonlocality in neural network synchrony, but do so to demonstrate the impossibility of approaching (at least some) large complex systems via network motifs by "network motifs". ZLS mechanisms, global network topology, and inter-network connectivity determine local network parameters and intra-network topology. This is why graph theory can provide a more useful (at least for examining functionality and dynamical features of the brain) approach than neural/neuronal network models or even coordinated neural population models. The application of classical physics (in the form of neurochemical electrophysiology) as a model of neuronal dynamics (at least epistemically) is ill-equipped to deal with a system more complex than those that motivated the use of (epistemically) indeterministic statistical mechanics long before quantum (ontological) indeterminacy

The development of modularity within the cognitive/brain sciences post-dated the development of "chaos theory" (in it's various incarnations, such as catastrophe theory). Work in complex systems has increased our understanding of the ways in which (at least for practical reasons) reductionism is often fundamentally inadequate, yet those dealing with one of the most complex systems known to exist are too often using a 19th century paradigm. While many researchers within the physical and life sciences have constructed new methods to treat complex systems as ore than the sum of its parts, those who deal with something as complex as the brain seem to be on average the most hardcore proponents of an epistemology that was motivated (and not intentionally developed) by successes within physics from Newton &Leibniz to Einstein that has no foundation whatsoever. The Hodgkin & Huxley equations are wrong. They are necessarily wrong, as they belong to classical physics. This is not to say that we require a new physics (as some systems biologists suggest) or some application of quantum mechanical mathematical structure (or ontology), but rather merely to note that the very foundations of the current approach are based and intricately tied to theories rejected by everybody.

Far more important (in my view) than quantum physics and whether even the statistical structure is useful here is the development of our mathematical methods used in no small way to demonstrate their own inadequacy (locally, everything is linear, but the higher the set dimensionality, the more arbitrarily small departures from a point are significant). It has been suggested that intra-network coordination is not only a key that connecting disparate brain regions allowing for a domain-general representation of even abstract knowledge (and the ways in which the perceptual systems are playing much more of a role in conceptual processing than classical cog. sci. supposes), but also that it is amodal, nonlocal network connectivity that binds perceptual-conceptual wholes. Simplifying brain regions into visual pathways and functional mediators of some neurobiological network is essential, but only if we are prepared to recognize the level of simplification involved (part of which, of course, involves knowing that we do not actually know how simplified such approaches are).

Cliffnote version: the ways in which neural regions function as a global workspace (insert sensationalist jargon here: nonlocality, emergence, supervenience, etc.), rather than compartmentalized, modular, and interconnected components seem to merit the use of models equipped to deal with such dynamical systems. Models of network connectivity are still typically modular, and a fair amount of evidence suggests they are inappropriately so.

Andrew: "Models of network connectivity are still typically modular, and a fair amount of evidence suggests they are inappropriately so."

On the contrary, it seems to me that the weight of evidence suggests that there *are* interconnected functional modules of specialized neuronal mechanisms in the human brain. How else can we explain such phenomena as size constancy, the moon illusion, the SMTT hallucination, etc.? See, for example, these papers: "Space, self, and the theater of consciousness", and "Where Am I? Redux" on my RG page.

I suppose there are a few ways to attempt to respond to such challenges to the views I expressed. The easiest would perhaps to say that "evidence suggests they are inappropriately so" rather than "evidence suggests this view is wrong" was intended to indicate that I am not asserting there exists no domain specificity nor that there are no specialized modular neural regions.

A stronger reply, perhaps, would be to assert that we haven't explained size constancy or the moon illusion (or at least that if we have, not all specialists in visual cognition would agree we have nor does the literature reflect that we have).

As is usually the case (from Piaget through Chomsky to modern developmental cognition), we find a trend of near-consensus on some aspect or component of developmental cognition during periods before titans like Renee Baillargeon & Elizabeth Spelke transformed the field. Their design of experimental paradigms not only demonstrated previous theories were inaccurate, but that these depended upon the incapacity of the subjects to indicate e.g., object permanence.

Likewise, we find despite numerous previous studies demonstrating that toddlers displayed size constancy, such conclusions have been challenged. For example, -

Granrud, C. E., & Schmechel, T. T. (2006). Development of size constancy in children: a test of the proximal mode sensitivity hypothesis. Perception & psychophysics, 68(8), 1372-1381.

Far more important, however, are studies such as this one (selected because I can link to the full text rather than depend upon others' access to ScienceDirect, Wiley Online, etc.):

Shape constancy, not size constancy: A (partial) explanation for the Müller-Lyer illusion

(http://csjarchive.cogsci.rpi.edu/proceedings/2009/papers/103/paper103.pdf)

As size-constancy has, in part, been an interesting area of research because of shape variance, studies like the above represent something of a challenge for the entirety of any conclusions based upon accepting size-constancy (which is not to say that it is an unanswerable challenge by any means).

More important still, at least from your perspective (I'd imagine), is how retinoid theory explains this phenomenon. This too faces challenges (again, by no means unanswerable) from empirical findings such as in e.g.,

Observer Movement and Size Constancy

(http://wexler.free.fr/papers/sico.pdf)

or

Higashiyama, A., & Adachi, K. (2006). Perceived size and perceived distance of targets viewed from between the legs: Evidence for proprioceptive theory. Vision research, 46(23), 3961-3976.

Finally, granted the phenomenon exists as classically described, and even that a model can account for it, there is the issue of whether the model actually does. Take the following study:

Sperandio, I., Chouinard, P. A., & Goodale, M. A. (2012). Retinotopic activity in V1 reflects the perceived and not the retinal size of an afterimage. Nature neuroscience, 15(4), 540-542.

I chose this one deliberately because I believe it largely either supports or at least is compatible with what I've read of your model. However, whereas you go into great detail on your model's computationally-validated success as well as its concordance with known dynamics the visual system's lower level components, I confess I'm not sure how to relate such work to aspects I find within studies like that above. How might one put in context a statement such as "[t]he retinoid system registers information in visual space and projects afferents to higher visual centers" or your neuronal network model with e.g.,

"our experiments suggest that retinal signals reaching V1 are modulated by distance cues in a manner that reflects the operation of size-constancy mechanisms in the real world. Namely, activity in V1 becomes more and more eccentric as the perceived size of an afterimage, or a real stimulus, increases (at least over viewing distances of 1.5 m) even though the size of the retinal image remains the same." (form the paper cited above)

The focus on cortical dynamics in the above reflects what I see as a larger trend. For example, in the nearly 2,000 pages of the 2-volume The Visual Neurosciences (among other more specific monographs, volumes, etc. than such a comprehensive reference source) I find almost no use of size constancy other than to indicate that it is likely going the way of "labeled lines" as the coordination required has proven to be more and more complex. I'm actually much more used to the term "perceptual constancy" as even when depth, color, size, and other (non)constancies are discussed I have found them approached holistically. None appear to be actually constant (not wholly), all appear to relate in numerous ways, and would-be neural amodal and/or multisensory "modules" have been implicated in visual processing for the last 20 years (and the visual system implicated in non-visual tasks, including abstract cognitive processes that 30 years ago everybody believed were amodal; now there exists the embodied cognition camp).

Finally, once again I fail to see the connection between computational discrimination, whether achieved by a neural network model or fuzzy discriminant analysis, and perceptual classification according to and structured by conceptual categorization. Habituation and adaptation of the type any neural network model, single cell, or algae are capable of does not become "learning" in the way that mice, cats, humans, chimps, etc. can simply because we design programs and hardware that can appear to identify faces, words, etc. but use the same mechanisms amoebas do. Syntactical manipulation, be it of a living system like a plant or a massively parallel state-of-the-art A.I. technology remains syntactical, not semantic. Since Turing, every decade has seen "the breakthrough"- computers themselves, the combinatorial/algorithmic transformational generative grammar, object-oriented programming, connectionism and neural networks, etc. But instead of receiving such breakthroughs, we find Minsky deriding the annual Loebner Turing test challenge, papers like Hodges' "Beyond the Turing Test" (Science) and French's "The Turing Test: The first 50 years", the applied A.I/computational intelligence/soft computing community (i.e., the one's writing software for auto-complete in smart phones or movie ad generators for Google) giving up on generative grammars in exchange for carefully annotated lexical databases & statistical machine learning, and the novelties we do see in the field are matched by novel or updated works like "What Computers Still Can't Do".

I think, perhaps, the question isn't "how can we explain..." the phenomena you mention, but rather "have we done so and if we have how have we shown this?" Having model of consciousness, whether it involves strange loops and algorithms upon algorithms or some unscientific metaphysical dualism maybe useful for discussion and to guide/direct research, but they are not explanations simply because they explain. They are explanations when we can show not only that they are plausible (e.g., when a neuronal model does what it is designed to do), but that they are actually the mechanisms of the brain. As much as I'd like to take a leaf out of the book of modern theoretical physics and cosmology and design untested models I adhere to for aesthetic reasons (mathematical elegance), that's not neuroscience (and for many, e.g., Davies, it isn't science either).

Dr Hackett,

Let us imagine two slightly different scenarios.

In the first scenario we have somehow been able to collate the entire experiences of an individual on a set of photographs. The subject's ability to recognize motor vehicles would seem to be a clear example of categorization. We would be able to extract all of the experiences of vehicles and we would note that they were all slightly different and yet the 'mind' had been able to abstract certain features from the collection of experiences to form the cognitive category.

Now, let us take the same individual and extract all of his/her experiences of Ayers Rock. Just is in the scenario above the individual has been able to abstract the essential features of Ayers Rock from a set of slightly different experiences. In this case we are unlikely to describe the cognitive ability to recognize Ayers Rock as an example of category formation.

The point is, having identified category formation as a significant feature of our cognitive lives it is also important to identify if it arises as a direct consequence of the brain's fundamental development processes.

Are we able to come to any consensus to the following question: when a human being navigates his or her daily lives they meet an enormous array of sensory stimuli. In order to make our navigation of the world possible we tend to categorise events into superordinate classes of events or categories. Early philosophers seemed to concentrate upon noun words as categories but obviously this expanded to all semantic types. What is the process that accounts for the formation of categories and the allocation of items to categories? Is this accomplished in a specific region or does this happen in different regions dependent upon sensory data type, data source, etc?

The medium is the message!

I suggest that how information is characterized is not always strictly a neurological problem. The Fractal Catalytic Model treats the brain as a catalyst that mediates transitions via 'fixed points' or points of invariance. The Fractal Catalytic Model treats language as simply another environmental stimulus. However, utterances embody an implicit order not only in terms of repetition and grammatical construction but also in terms of the relationship between how language is used in the world(occurence of words, etc). In this respect, words and structures embody an implicit ontology that the brain makes explicit as conceptualization.

I suggest that how this process structures the brain is, to a significant degree, determined by the structures implicit in language use rather than as the result of some sort of predetermined chomskyesque centre for language processing.

Andrew: "Finally, granted the phenomenon [size constancy] exists as classically described, and even that a model can account for it, there is the issue of whether the model actually does. Take the following study:

Sperandio, I., Chouinard, P. A., & Goodale, M. A. (2012). Retinotopic activity in V1 reflects the perceived and not the retinal size of an afterimage. Nature neuroscience, 15(4), 540-542....I chose this one deliberately because I believe it largely either supports or at least is compatible with what I've read of your model. However, whereas you go into great detail on your model's computationally-validated success as well as its concordance with known dynamics the visual system's lower level components, I confess I'm not sure how to relate such work to aspects I find within studies like that above. How might one put in context a statement such as "[t]he retinoid system registers information in visual space and projects afferents to higher visual centers" or your neuronal network model with e.g.,..."our experiments suggest that retinal signals reaching V1 are modulated by distance cues in a manner that reflects the operation of size-constancy mechanisms in the real world. Namely, activity in V1 becomes more and more eccentric as the perceived size of an afterimage, or a real stimulus, increases (at least over viewing distances of 1.5 m) even though the size of the retinal image remains the same." (form the paper cited above)".

Look at *Projections from the 3-D Retinoid: Size Constancy* on pp. 89 - 93 in "Accessory Circuits" on my RG page. In particular, Fig. 5.8 shows the kind of neuronal structure that clearly explains why "... activity in V1 becomes more and more eccentric as the perceived size of an afterimage, or a real stimulus, increases (at least over viewing distances of 1.5 m) even though the size of the retinal image remains the same." (from Sperandio et al, *Nature Neuroscience*, 2012).

Andrew: "I think, perhaps, the question isn't "how can we explain..." the phenomena you mention, but rather "have we done so and if we have how have we shown this?"

I agree that if an explanatory model is proposed it should be empirically tested for its validity. In science we take the successful prediction of previously inexplicable phenomena as the gold standard for the validity of a theoretical model. As far as I can see, the progressive diminution of the perceived size of the moon as it rises above the horizon, and the extraordinary demonstration in the SMTT experiment of a vivid visual hallucination that can be systematically controlled by the person experiencing the hallucination, are two powerful empirical confirmations of the retinoid model.

By the way, the findings of Baillargeon and Spelke provide further evidence in support of the retinoid model.

Regarding Dr. Hackett's question on consensus-

The answer is yes: there is a consensus on the question(s) concerning categorization. Unfortunately, the consensus position is that their is no consensus position view As Dr. Antrobus remarked, neuroscience is in its infancy. I would add that unlike a number of fields also in their infancy, neuroscience is has strong ties to older fields in ways that present more challenges than, perhaps, other burgeoning fields. Memory provides a simple and interesting example, as the *categories* of memory types come from early work in cognitive science (Miller's famous "The Magical Number 7" which perhaps my favorite opening line of any journal article), functional neuroimaging, behavioral studies on rodents and humans [insert lawyer joke here], neurophysiology, etc. As a result, memory "types" overlap, have at best questionable empirical support, are used inconsistently in the literature, and possible side-effects of studying them include headache, nausea, memory loss, and career changes.

This does not mean theories are all over the place. For many fundamental questions, such as whether a category is specific to some brain region or distributed across regions, whether or not concepts are amodal (and relatedly whether or not cognition is embodied), everybody falls more or less into one of two camps. Also, these tend to be related: a neuroscientist who believes cognition is embodied will believe categories are distributed across brain regions.

Other questions are not so evenly divided. These include:

1) How "broad" the neural representation of a category is. When I see a dog, there exists some neural representation (distributed across brain regions or not) that I rely on to categorize this visual stimulus as a "dog". But do I have a dedicated neural representation of the category "dog"? Or do I rely on a more general neural representation such as the category "animal"?

2) Regardless of where neural representations of however specific or broad a category are, what mechanisms are behind categorization (e.g., how salient is shape vs. color and to what extent is this dependent upon context, the category in question, both category and context, etc.)?

Because we are not omniscient, and because proof is found only in formal logical models, no proposed theoretical model is beyond question and criticism. Science has never been without internal disputes. Progress in science has often required a departure from consensus.

Quite! Today's theories and models are the best representations of what we have observed so far. But they are simply stepping stones to better theories and models. In today's research arenas, their primary value is to be replaced or modified by better models within a few months or a year.

Dr. Trehub-

First, thank you for the help locating the relation between your model and the quote I had requested. It was much appreciated. I would request that, if in the future, you are kind enough to point me to sections of your book, you simply give the page numbers and not the chapter title, because

1) Even though I own the book I didn’t recognize the chapter name and went off looking for a paper

2) Although your entire book is apparently available for free, I very much prefer hard copies and in this case I have one already.

I'm still not quite certain about a number of things. Part of this seems to be different orientations and experience, so I appreciate your patience with me. While much of the visual system is familiar to me once we get to the brain, I know more about electrodynamics than I do the biophysics of fovea.

I’m also used to modelling using NEURON and MATLAB, and “sparse coding” brings to mind SparseLab not the neurodynamics of facial recognition. However, what is most intriguing about your model is how much it uses the language of neuroscience and the structure, method, and mathematics of soft computing/machine learning. Of course plenty of work in computational neuroscience makes little attempt to model the neural systems in questions or any system, as in e.g.,

Ma, L. (2011). Retinotopic Sparse Representation of Natural Images

Better yet, take:

Ouyang, Y., & Sang, N. (2013). “A facial expression recognition method by fusing multiple sparse representation based classifiers” (ISNN 2013 vol. I)

This isn’t a model, but a method or technique. I know what to expect: the required inclusion of SOMETHING “Bayesian”, k-means, nearest neighbor, feature extraction, even a confusion matrix as found in your contribution to Quantitative Analyses of Behavior-

“A confusion matrix for Hand-Print Alphabetic Characters: Testing a Neuronal Model”.

These and other classification and pattern recognition methods are similar to yours in some ways yet none are neuronal models. Your neuronal model is clearly identifiable as such through e.g., the many references to ESPS, axons & dendrites, synaptic junctions, retinal input, etc. However, change the terms used and we find a somewhat older approach to pattern recognition, not a neuronal model. Instead of e.g., some model OF neuronal patterns (e.g., spatio-temporal firing patterns), we find a pattern recognition method. You clearly relate the technique in theoretic ways to neuronal processes, just not in ways I’m used to seeing (i.e., the equations for the physical system’s actual properties/processes/evolution and how the models’ results match e.g., neuroimaging data).

I hope you’ll bear with me, as rather central to both this topic & what I have asserted (rightly or wrongly) is the distinction between algorithms capable of something like procedural learning but not categorization. I know plenty of classification algorithms and other increasingly sophisticated ways we can make computers take input the way a calculator does and return output. What I don’t know of is anybody who has created a program or system capable of any representation of a conceptual network through which perceptual input is categorized: “Although many of today’s connectionists and computational neuroscientists emphasize the explanatory role of association, many of them also combine association with other explanatory constructs, as per weak associationism (cf...The Cognitive Brain; pp. 243–245...). What remains to be determined is which neural networks, organized in what way, actually explain cognition and which role association and other explanatory constructs should play in a theory of cognition.”

Piccinini, G., & Scarantino, A. (2011). Information processing, computation, and cognition.

The pages you referred me to do indeed provide a clear explanation for what I quoted, but I specifically chose that paper because it was in agreement with you (as I said) yet I could not readily relate your model to it. Obviously, I was unclear, so I’ll try to another method:

Barnes, T., & Mingolla, E. (2013). A neural model of visual figure-ground segregation from kinetic occlusion.

The above is an neural model and we do find schematic representations, learning sequences, etc. But the model is designed to replicate the neurodynamics through equations grounded not just in what they can accomplish but in what is known about the make-up of the system responsible for the phenomenon the model is tested against. In contrast, looking at your model (as useful as it may be in multiple ways), I can’t help but wonder why a classification method uses biological labels for mathematical notations that are clearly not biologically plausible. Hence the question (hopefully more clearly articulated): how can your model “clearly explain” what’s going on in neural regions not mentioned through a mathematical model of a mechanism that is not related to any biophysical parameters or equations of state; isn’t able to replicate either human categorization or modern pattern recognition, classification, and clustering methods (among other similar techniques and combinations of these); and is theoretically grounded in an interpretation of certain phenomena (and not others) but not in relation to the dynamics of the physical system in question?

The retinoid model may very well be absolutely the path we should be taking and the best framework available, but my focus isn’t the visual system (except insofar as conceptualization requires perception and perceptual experiences) and I cannot see how your model can explain cognitive processes I do focus, such as language and the ways in which conceptual networks are constructed by and used for categorization. So I am trying to understand the disconnects (and, clearly, not too well).

For example, and staying within the realm of vision, are there not other phenomena than the several you mention?

“Once upon a time it was widely believed that human observers built up a complete representation of everything in their visual field. More precisely, it was believed that the stable and richly-detailed world viewed by a normal observer gave rise to a stable and richly-detailed representation within, a ‘picture’ that could be used for all subsequent visual and visuomotor operations (e.g... Trehub, 1991)...With such a visual system, presumably, the observer could live happily ever after.

This idea of a stable, general-purpose internal picture accords nicely with subjective experience. It does not, however, accord quite so nicely with objective fact… recent studies indicate that perception itself is susceptible to various forms of induced blindness: repetition blindness…inattentional…change blindness…and an attentional blink”

Rensink, R. A. (2000). Seeing, sensing, and scrutinizing.

You write “If we are to understand the biophysical basis of self, then we must first give a plausible neuronal account of how the human brain is able to represent an object-filled space from an egocentric perspective”, yet we find empirical evidence this is a not so visually-based a process as it seems you suggest:

“Through sensory-motor processing, all these frames become aligned in a representation of the common 3D space they occupy. The right panel shows a similar situation, but with a critical difference: Instead of the image of the target being displayed in its source location, it is displaced for viewing to a remote screen. The spatial frame of reference for the visible image is arbitrarily translated and rotated relative to the frames representing the action and the word. In effect, the operator’s eye has become ‘‘disembodied”.”

Klatzky, R. L., Wu, B., & Stetten, G. (2010). The disembodied eye: Consequences of displacing perception from action.

Or, looking at the retinoid model as a basis for egocentric perspective, given its architecture is founded on vision, how might it be related to one or more of the following positions:

“Nowadays, the vital importance of space perception in the relationship between perception and action is recognized and discussed in three largely independent research camps: (1) the ecological camp, (2) the two-visual-systems camp, and (3) the constructivist camp. The theoretical points of view emphasized in these camps are not mutually exclusive. In fact, by emphasizing (1) light, (2) brain, and (3) behaviour, they neatly complement each other”

(from the intro remarks to the special issue of Visual Cognition (II 2/3) Visual Space Perception and Action).

Do the same mechanisms behind the visual phenomena you identify as key underlie the ways in which manipulation of multisensory processes can distort the fundamental “egocentric perspective” you refer to? (see e.g., The Neural Basis of Multisensory Processes and in particular the papers “Peripersonal Space: A Multisensory Interface for Body–Object Interactions” & “Multisensory Perception and Bodily Self-Consciousness From Out-of-Body to Inside- Body Experience”)

For me the visual system is part of the perceptual system and relates to categorization through conceptual networks, domain-general embodied networks or massively-modular and conceptually amodal. A model that is not merely an extension of one theoretical way in which a single sensory modality functions, but quite fundamentally based on neuronal activity outside of the brain that extends the kind of classical receptive field models and their focus on discrimination to the entirety of categorization (from sensory-motor to perceptual-conceptual) without either a model of the physical system itself or for most of research in cognitive linguistics over the past 10+ years is, when one thinks about it, no more or less ambitious than most accounts of consciousness and cognition, but only (at least to I feel) if understood very generally. However, you don’t appear to think of it as such.

Perhaps this isn’t the forum for such a discussion, however. For one thing, my original drafted response was almost 10x the length of this one. For another, I feel I my statements are challenging your view when that is not my desire or intent, yet I can’t think of how to ask the questions I have without identifying the problems that motivate them.

Andrew: "In contrast, looking at your model (as useful as it may be in multiple ways), I can’t help but wonder why a classification method uses biological labels for mathematical notations that are clearly not biologically plausible."

It seems to me that you have it backwards here. I use mathematical concepts (e.g., integration, ordinal logic, etc.) to describe biological processes (e.g., cumulative increase in post-synaptic potential, ordered latency to axonal discharge, etc.), and these are clearly biologically plausible.

Andrew: " A model that is not merely an extension of one theoretical way in which a single sensory modality functions, but quite fundamentally based on neuronal activity outside of the brain that extends the kind of classical receptive field models and their focus on discrimination to the entirety of categorization (from sensory-motor to perceptual-conceptual) without either a model of the physical system itself or for most of research in cognitive linguistics over the past 10+ years is, when one thinks about it, no more or less ambitious than most accounts of consciousness and cognition, but only (at least to I feel) if understood very generally. However, you don’t appear to think of it as such."

I have a problem in understanding the point of your above sentence. Can you express it more directly?

As a general comment, I would say that each investigator might have an intuition about the best theoretical model for explaining human knowledge of the world, but in science, evidence trumps intuition. So the candidate model that is best able to successfully explain/predict relevant evidence eventually becomes the standard.

For more evidence in support of the retinoid model see "Where Am I? Redux", and for a summary of what I think is the right approach to the study of consciousness and cognition see "A Foundation for the Scientific Study of Consciousness". Both can be found on my RG page.

This thread asks about the neurological basis for category formation, but I think the subtext really concerns the neurological basis for human *understanding* of the world. I argue that no theoretical model that purports to explain human understanding of the world is well founded unless it accounts for subjectivity.

Dr. Trehub-

Thanks for the response and once again for your patience. I believe I can address both your request for clarification and what I meant by “biologically plausible” at once (and perhaps resolve a larger issue). As I stated before, we’re not only coming at this from different perspectives, but also different levels of experience (yours vastly exceeds my own) and different backgrounds. Several issues I have stem from (I think) differences relating to models and modelling. You stated:

“I use mathematical concepts…to describe biological processes…and these are clearly biologically plausible”

I think the following is pretty uncontroversial: "A crucial issue in any computational model is what the appropriate level of abstraction is. Although in theory we could dissect and model each neuron at the smallest level of detail allowed by current technology...in practice only a few measurements of the system parameters would be available at that level...Superfluous detail can also make it difficult to understand the model and to generate predictions based on it. Fortunately, such detailed simulations are often unnecessary for understanding high-level behavior" (3.1 of Computational Maps in the Visual Cortex by Miikkulainen, Bednar, Choe, & Sirosh).

“The A17 Paradox” by Diamond & Grimes (in The Computing Dendrite: From Structure to Function) is, I think, doubly useful. First, it nicely ilustrates the trade-offs in a model’s level of abstraction. Only at a very low level could we find that “A17 amacrine cells appear to forsake typical dendritic integration (i.e., collecting and combining synaptic inputs to form a single output signal that is distributed via the axon) in favor of distributed, parallel processing”. Despite the detailed level of the model, it informs us of much more abstract, higher level models. Many a multi-layered neural network model functions as a unit via mechanisms (summing input as a whole with a single output) that are simplistic compared to “the retinal A17 amacrine cell, in which morphological and biophysical specializations create hundreds of highly compartmentalized, functionally independent computational units within their dendrites”. On the other hand, to my knowledge nobody is using A17 cellular characteristics to explore learning and classification mechanisms.

Generally, the higher the abstraction, the more the descriptions, equations, and terminology reflect this. In the above paper, everything is low-level, everything describes the specific cells and how they function in their specific contexts of the visual system. Your paper “Neuronal models for cognitive processes: Networks for learning, perception and imagination” begins (more or less) at this level. Rather than a schematic figure depicting some high-level cognitive process, we find a picture (Fig. 1) of a biological synaptic junction. Much is similar to what we find in chapter 2 of your book over a decade later. Yet although even then you mention that there is “no physiological confirmation of the specific details” for your model of synaptic plasticity, your earlier paper goes from a proposed retinal pattern recognition mechanism to your retinoid model (which is the foundation for your book as well as consciousness, as in e.g., “Space, self, and the theater of consciousness”). Yet so basic, ubiquitous, and fundamental a component of neural networks as is a threshold nonetheless is questionable at best (especially represented by a single parameter extrapolated to any and all neural networks), given that

1) Neurons have no well-defined thresholds (though they may have a well-defined rheobase)

2) The main classes of neuron types, integrators and resonators, are not only quite broad but also somewhat misleading given that e.g., a neuron’s characteristically integrating behavior need not remain so, as such a neuron can resonate

3) Resonators may not have even a well-defined rheobase

4) Despite the ongoing debate over the neural code, it is probably safe to say that whatever their contribution to neural information may be, spiking is not dependent upon any threshold so much as by a number of dynamic parameters governed by everything from the “classes” of firing patterns one can divide neurons into, dendritic structure, neuronal type, and whether the neuron even has a well-defined rheobase.

The issue of biological plausibility I had in mind was when one uses the kinds of terms we might find in a very low-level model in a high-level model instead. As models move away from neurobiological models or models of neurophysiological processes we increasingly see terms specific to neurobiology dropped. For example, the following is a biologically inspired neural network, yet it is described as artificial, as an algorithm (not a model), and is STILL limited in scope:

http://ijcai.org/papers13/Papers/IJCAI13-185.pdf

Contrast such an approach to one that seeks to be biologically plausible:

http://www.columbia.edu/cu/biology/faculty/yuste/Publications/yuste.neuron11.pdf

http://www.sciencedirect.com/science/article/pii/S0896627303008389

Your model type is more or this kind:

http://www.neurotheory.columbia.edu/~dmarti/publications/marti_rinzel_neco2012.pdf

or better yet

http://redwood.berkeley.edu/fsommer/papers/KnoblauchPalmSommer09preprint.pdf

in that plausibility or utility is based far more on what the model can DO, rather than being based on, built upon, and tested against what we know about the dynamics of a neuron or neuronal population within some particular brain region. Yet we do not find the same kind of neurobiological language. So I guess by “biologically plausible” I mean partly what you yourself have described (the lack of “direct physiological confirmation”) for the basis of your model, which can only be said to be vastly increased once it is abstracted from a model of synaptic plasticity to basically all cognition and more (including consciousness). As we know that neuronal networks differ qualitatively, and that many features you include at the lower-level (like threshold values and your synaptic transfer weight) are simplifications even at that level ("all models are wrong" after all, but "some are useful" so I don't mean that as a criticism in the slightest but as a given regardless of the model), it seems strange to me to see lower-level terminology in what is basically an algorithm for character recognition (at least in that paper’s context; obviously as a whole the retinoid model is far more).

As for evidence, not only do we find models for language that are multisensory, embodied, etc., but are based not on an ability to learn algorithmically or classify patterns but functional neuroimaging and similar tests of the human cognitive-perceptual system:

http://asociatiaromanadehipnoza.ro/wp-content/uploads/2013/08/Active-perception-sensorimotor.pdf

Somatotopic Representation of Action Words in Human Motor and Premotor Cortex

http://www.informatik.uni-hamburg.de/WTM/publications/

http://www.researchgate.net/publication/51173938_Conceptual_representations_in_mind_and_brain_theoretical_developments_current_evidence_and_future_directions/file/e0b495268e562ad7a7.pdf

Embodied Language Understanding with a Multiple Timescale Recurrent Neural Network

etc.

I'm not endorsing any of the above and I agree that this thread seems to be more about how we understand the world, but I do not think that we can safely say an algorithmic/machine learning classification explanation is our best approach/framework, and even if it is, that the retinoid model is the only one that should be looked at. Even if you are absolutely correct, given the nature of the thread, is it not better to respond to a question so broad concerning so many fields and even more unknowns that we give not “the” answer, but point rather to what kinds of answers there are (and where)?

Article Conceptual representations in mind and brain: Theoretical de...

I recommend the article "How Much Does a Shared Name Make Things Similar? Linguistic Labels, Similarity, and the Development of Inductive Inference" Sloutsky et al. (2001). If you want a quick run down of the paper then I have presentation slides for the paper which you may find useful at: http://sciepsy.wordpress.com/2014/02/23/the-development-of-categorization/

http://www.psy.cmu.edu/~cognitivedevelopmentlab/Sloutsky-Lo-Fisher-2001.pdf

Many thanks Jenna.

This is a very important topic. I intend to try to read the preceding answers before commenting

It’s a kind of a dilemma. How to investigate the brain mechanisms of a phenomenon as categorization? While the “full range” of cognitive-phenomenological approaches is limited to humans, any investigation of animal nervous systems physiology deals with the problem of defining the cognitive aspect of the learning process.

Work that has captured the multidimensionality of categorization formation on the level of neuronal circuits by using advanced neurophysiology techniques combined with neurodynamic mathematical approaches is from Ohl et al., 2001; Nature

(http://www.nature.com/nature/journal/v412/n6848/full/412733a0.html ).

An introduction and overview of research around the neuronal basis of categorization is given by the authors here:

http://sulcus.berkeley.edu/wjf/BG_Ohl%20et%20al.%20AuditNeurodyn.pdf

In a recent paper (Gainotti G. The organization and dissolution of semantic-conceptual knowledge: Is the 'amodal hub' the only plausible model? Brain and Cognition 75, 299-309, 2011) I gave the following answer to this question:

"More recently, some authors have offered a more specific account of the memory functions subsumed by the perirhinal cortex, hypothesizing that it may give an important contribution to semantic memory (Davies et al., 2004) and contribute to the construction of stimulus familiarity. In particular, Rolls et al. (Rolls, 2000; Rolls and Deco, 2002; Rolls et al., 2005) have provided an increasing amount of neurophysiological data, showing that the perirhinal cortex implements a long-term form of familiarity memory, which reflects the gradual build-up of neuronal responses over several hundred presentations of a stimulus and may, thus, represent the degree of long-term familiarity of stimuli. The convergence in the perirhinal cortex of highly processed visual data with other kinds of perceptual information and the acquisition of progressive familiarity feelings with the corresponding classes of objects might, therefore strongly contribute to the construction of the ‘sensory information based semantic system’. This is, in my opinion, a possible neurological basis for the simplest forms of category formation

Filippo, I wonder if the sensory level processing really won't provide the 'finest' level of classification, which then becomes coarser when aggregated with data from other sense processors into object classification in the scenario you suggest?

I'm remembering a paper from a long time ago that dealt with the contributing electrical signals that are aggregated into the heartbeat waveform we all know and love. The heartbeat is a good and well-used descriptive but it's constitutional members are difficult to discern.

It seems to me that you have to specify whether you are classifying objects or *features* of objects. The feature of shape alone is often sufficient to classify an object[1], but if we want to classify different kinds of roses, for example, we may need to classify on the basis of both shape and color. The fact that shape and color are detected in different parts of the visual system raises the very difficult problem of how different sensory features, represented in different parts of the brain, are properly bound together in our phenomenal experience. How, for example, is the color of a red rose represented only within the shape contours of the rose and not in its stem or the vase holding it? A neuronal mechanism that can accomplish this kind of spatio-temporal binding is described in *The Cognitive Brain*[2].

[1] "Sparse Coding of Faces in a Neuronal Model ...", (on my RG page).

[2] "Analysis and Representation of Object Relations", (on my RG page.

In the neuroscience and cognitive psychology literature (e.g. Damasio, A.R., 1989, Cognition 33, 25-62.; Damasio, A.R., 1990, Trends Neurosci. 13, 95-98.) it has been assumed that the representation of objects (and of natural categories) requires the convergence of features in the representation of whole objects. The functional architecture of the system proposed by Damasio (1989 and 1990) is constituted by: 1. neuron ensembles in the ‘early’ sensory and motor cortices; 2. neuron ensembles located downstream from the former throughout single modalities cortices (‘local convergence zones’); 3. neuron ensembles located downstream from the previous ones, throughout higher-order association cortices (‘ higher-order convergence zones’). Furthermore, as different sensory, motor and verbal modalities play a different role in the organization of different categories (e.g. - visual shape and motion for animals, - visual shape, color, hodor and taste for vegetables and - shape, action movement and tactile inputs for artifacts), the cortical areas underlying these categories will correspond to the areas of convergence of the sources of knowledge playing a critical role in each category.

Anatomo-clinical and experimental data are consistent with this model.

Guido: "The functional architecture of the system proposed by Damasio (1989 and 1990) is constituted by: 1. neuron ensembles in the ‘early’ sensory and motor cortices; 2. neuron ensembles located downstream from the former throughout single modalities cortices (‘local convergence zones’); 3. neuron ensembles located downstream from the previous ones, throughout higher-order association cortices (‘ higher-order convergence zones’)."

Unfortunately, this kind of functional anatomical description is a *black box system* like those commonly proposed as explanatory models in neuroscience. But it really avoids the essential questions. What are the neuronal mechanisms inside of the anatomically located black boxes/neuron ensembles, and how do they do the job that they are supposed to do? Calling the necessary mechanisms "neuron ensembles", tells us nothing about how they work.

Bringing all the different sensory features together in proper spatio-temporal register, within the brain's representation of the volumetric space around you, from your own privileged perspective (subjectivity), is what your cognitive brain is able to do. Talking about neuron ensembles doesn't get us very far in explaining this remarkable biological accomplishment.

As a lay person, it seems to me that the formation of categories begins with associations of stimuli with similar previously stored neural states. If the association is "close enough" to a previous state, the stimuli are accepted as belonging to the previous state category. However, if predictions or other characteristics of the assembly of the new stimuli differ sufficiently from those of the old category, then perhaps a totally new category is called for, or a refinement of the old category is needed.

Arnold, I agree that the formation of categories at the brain level is a highly complex phenomenon, which probably requires a cascade of related mechanisms and that the Damasio's model of the 'higher-order convergence zones' is only a component of this system. A more basic integrative mechanism could reside in the correlation learning theory, proposed by Hebb (1949), which aims to identify the neural mechanisms through which specific cortical representations are learned. This model assumes that the perceptual experience of objects depends on neural activity in multiple regions activated simultaneously and that neurons which fire together for a while strengthen their mutual connections and become more tightly associated, developing a functional unit labelled ‘cell assembly’. The advantage of the Damasio's model is that a hierarchical organization is necessary to pass from the representation of 'objects' to the representation of categories and that the model of the 'higher-order convergence zones' provides this hierarchical organization. Furthermore, since the weight of different 'sources of knowledge' is different for different categories and that each sensory or motor information is represented in specific cortical areas, it should be possible to predict the the anatomical location of the 'higher-order convergence zones' (cortical areas/networks) which play a critical role in the development of different categories. Empirical clinical and experimental data are consistent with these predictions

Guido: "A more basic integrative mechanism could reside in the correlation learning theory, proposed by Hebb (1949), which aims to identify the neural mechanisms through which specific cortical representations are learned."

Hebb's original proposal has to be modified to account for effective learning in our cognitive systems. *The Cognitive Brain* (MIT Press 1991) presents a large scale neuronal model that explains the formation of categories and much more about the mechanisms of cognition. To get an idea of the details, see the following publications on my RG page:

"Learning, Imagery, Tokens, and Types: The Synaptic Matrix"

"Overview and Reflections"

"Space, self, and the theater of consciousness"

"Where Am I? Redux"

Check out Jack Gallant's group's recent work on mapping semantics in the brain using the population receptive field methodology. This gives a profound empirical and detailed view into how our brain categorize. Repeating this kind of study with native speakers of other languages could provide some remarkable insights!

Huth AG, Nishimoto S, Vu AT, Gallant JL. A continuous semantic space describes the representation of thousands of object and action categories across the human brain. Neuron. 2012 Dec 20;76(6):1210-24.

Robert, many thanks for this link. The paper contains some excellent information. Do you work in categorisation?

One should not mistaken between classification and categorization. The former assumes existing category structure and determines the most likely category for a given (sensory) sample, while the latter builds up a category structure.

Neural-level bottom-up categorization could be obtained with multi-level unsupervised Hebbian learning in the so-called Deep Networks, exploiting co-occurency of those features in multiple sensory samples, gradually emerging accross depth increasingly more abstract sets of features, that at certaint point we might call "categories" (or at least would be close to popular categorization systems).

For a practical example, with indicative neural basis, you could check: Stoianov & Zorzi, "Emergence of a 'Visual Number Sense' in Hierarchical Generative models", Nature Neurosc. 15, 194–196 (2012). In that work, abstract numerical categories emerge from bottom-up processing of visual sensory information. This model explains animal neural-level data and human immaging- and behavioural-data.

For insights on these models and tutorial on using them you could check: Testolin, Stoianov, et al., (2013) Deep unsupervised learning on a desktop PC: a primer for cognitive scientists. Frontiers in Psychology 4:251.

Those who prefer a more abstract probabilistic-level of reasoning could check Sanborn & Griffiths, Psych Review 2010, pp 1144-1167. Those have been successfully applied in various higher-level domains, but requires specific technical knowledge.

Categorization certainly benefit from converging input from multiple domains.

Cognitive linguistics is quite clear, firstly, that categories are encyclopedic; this requires a system-wide approach. Secondly, philosophy suggests that approaches can be Cartesian or phenomenal in a Heideggerian sense; this implies a core default structure with a more flexible contextual backup oriented around relevancy. Thirdly, neonates cannot categorize but adults can – this implies that categorization must be integrated into childhood development. Finally, perceiving as an intermediate developmental stage (between apprehending and understanding) involves pattern-matching which is a kind of instantaneous categorization. Hegel puts it together in the initial sections of ‘Phenomenology of Mind.’ You’ll find a line-by-line commentary of these sections (book-length), with diagrams, at http://www.cognitivestyles.com in sections ‘8a Developing Visual Circuits’ to ‘8f Unhappy Consciousness’ (Hegel’s language) in the drop-down list. For a “hierarchical framework of categories” (genus-species-particular), which I take is the essence of your question, look in particular at the evolution of the circuit 1 Classification and its relation to Cartesian thought. As far as neurological correlates (“neuroscience research”) are concerned, you will find that the diagrams map onto the brain (see sections ‘1a Lateral Frontopolar’ to ‘1f Aspects of Depression’ in the drop-down list, again book-length).

More specifically, you mention the understanding of categories (“We do not experience the world as a series of unique events but UNDERSTAND experiences within a perhaps hierarchical framework of CATEGORIES.”). An answer will have to incorporate the following aspects, which I pose as a series of questions: How does understanding determine the genus in classification? Is it the only strategy which determines the genus? Why is the genus sometimes the thing which is missing in the object, and not the object itself (e.g. my car has a flat tire, I’m in the desert and I have no spare, making the genus of classification an operative good replacement tire, which is what is missing in my car and which is NOT IMMEDIATELY PRESENT as an object; or the scientist conceptualizes a new rocket engine which is NOT ANYWHERE PRESENT and works on that as his contribution to the manufacture of the rocket) – that is, the classification genus in the first instance is the tire and not the car of which it is a part, and in the second instance the engine and not the rocket of which it is a component? How does understanding then ‘realize’ the context of the missing component and incorporate it into the genus? How does it adapt other objects through an ‘as’ structure which views them for instance as a possible spare tire for my car and places them in the species node under the new genus of what is missing, using what is ‘realized’ as an aid? By extension, how does the species adjust so that the same receptacle (using another example) is placed at times under ‘heavy objects to throw in self-defense’ or ‘mug-like shape that can hold coffee’? How does the specific in classification become the foreground object within perceiving in such as way that the perceived in the external does not throw the mind outside of the genus in classification? What are the intermediate and deep backgrounds within perceiving and how are they formed in a way that is consistent with this flexibility?

More generally, how do the visual circuits form in childhood development and what role do they play in category classification which is simultaneously memory reconstruction? How is it that classification respects Newton’s Three Laws for moving objects, and moves forward in time through simulation so that we can for instance catch a moving ball that comes towards us? How is it that the geographer can look at interlocking circles on a map and actually see the physical mountains in her imagination? How can intersecting lines in a bubble chamber be chicken scratches to a layman and at the same time nuclear events to a trained physicist? What is a paradigm shift as it relates to classification and why is it impossible to ‘see’ and thus classify outside of one’s paradigm? These are some of the questions which are addressed and answered.

Just to get it out of the way: Newton's three laws are all wrong so I don't see why anything that "respects" them is somehow significant because it does. Also, to repeat what I said some time ago and what Dr. Stoianov said more recently, classification isn't categorization. For what it's worth, I don't know of a cognitive linguist who would disagree (Langacker, Lakoff, Talmy, Geeraerts, Dirven, Johnson, Fauconnier, Croft, Michaelis, Goldberg, Fillmore, Dancygier, Sweester, and other names who could be said to fall under that vast umbrella).

Of course, Dr. Stoianov's reason for distinguishing the two is not my own. I don't disagree, it's just that such a distinction is unimportant compared to the much bolder (and less substantiated) ones that I would make. I do not, for example, find anything particularly similar between classifiers in neuronal or neural network models (or machine learning, computational intelligence paradigms, soft computing, etc.) and the way humans classify. Computational models are able to sort input into "categories" but they cannot classify according to these because we cannot formalize concepts. Over the past few hundred years, some of the greatest minds have developed the central tool of the sciences: mathematics. The calculus, for example, may be credited to Newton and Leibniz, but it was not until Weierstraß that limits had a sufficiently formal, rigorous definition for analysis. Ambiguity is not tolerated in mathematics. Jean Dieudonné, in a French calculus textbook he wrote in the 80s, seemingly couldn't resist showing his disdain for many-valued functions. As John Hubbard pointed out in support, computers provide the perfect illustration that Dieudonné was right: they do not tolerate ambiguity. They are physical instantiations of an abstract algebra and no more capable of brain-like classification than a pocket calculator (which they are equivalent to). Catagories are conceptual, and computation is syntactic. Unless we figure out how to formalize that which is defined in opposition to formal and define definition (or give the meaning of meaning) we have nothing to build on to make the qualitative leap from sea slug learning of the type computers are capable of to the conceptual processing that brains are. 300 years removing semantics from mathematics and now we're using it to try and model meaning. But what else do we have to use?

It would be particularly instructive for one to spoil her hands with various types of models, at various levels (e.g., "neural networks", "probabilistic networks", etc.), to actually undestand how those methods could contribute to the undestanding of the functioning and structure of the biological neural systems.

Sure, we are talking of approximations, with great potential, though, to open the Pandora box. It's the integration of those methods with various other methods based on observation that is particularly benefitial. For example, when a certain computational model can explain a biological system at multiple levels, e.g., behaviour and single-neural level, then one could say with greater confidence that the abstracted mechanism corresponds at a given (still high) level to the biological system. Some are also interested in greater detail, e.g., functioning within a single neuron, and, e.g., spiking neural networks try to provide such an answer. Others go even deeper. Sure, the deeper you go, the greater the level of complication and the smaller the global function you can explore, mostly due to limits of computational resources. Large human-brain projects have the ambition to put together the depth and breadth of functioning explained.

Back to categorization, unsupervised learning of milions of samples can develop a representational system (based on just hundreds or thousends of features) capable to generate the experienced samples with reasonable quality, retaining main features (invariances) and throwing away useless ones (noise). These representations might not be exactly the same as the ones we know, but they could bring homology to known system of categories. Thus, these deep networks have the great potential to explain the functioning of the sensorial systems, by providing to deeper (e.g., language-processing) structures just essential category-like information.

I have the advantage on standing on the works and developments of better minds that came before me. McCulloch & Pitts, Hebb, Hopfield, Conway, etc. I've been following ICANN, HCII, ISNN, and similar symposiums and conferences for years. While others had to develop their programming environments I get to use MATLAB and NEURON (not to mention C++ with the addition of SymbolicC++ libraries, Java, etc.). But when I get paid to use gene expression or neural networks in consult work that are, in general, more sophisticated than models which seek to be biologically plausible (including my own), it does not inspire confidence. Google Translate doesn't use generative linguistics nor does DeepQA, and biologically inspired methods include swarm intelligence which is no more helpful in understanding neural systems than VSM or PCA.

Computational methods are incapable of the kind of conceptual representation that Tolman, Ritchie, & Kalish showed rats possess, but with it we can implement the mathematical models that McCulloch & Pitts developed in the 40s and the learning mechanisms that Eric Kandel won the Nobel prize for in 2000. That's why I call it sea slug learning. Because it quite literally is the implementation of the kind of learning his investigation of Aplysia californica was instrumental in furthering. Combining these with memory and the precision of computer architecture allows us to do amazing things, but I have yet to see it do something qualitatively different than what we have been doing for decades. Perhaps I'm just jaded. But I will be happy to be wrong!

As for capability of neural-network models, who knows why Google got Jeff Hinton, the key developer of the Deep Networks I mentioned before ... ? ( I nevertheless hope J Hinton will still make pubblic key novelties). And who knows why Google started to actually use those methods right after that ? It could be quite instructive and assuring to scepticists to know what is behind some of the recent speech recognition technology of the modern PAD devices. It is also very easy to predict what will lay behind the forthcomming image- and video-recognition technologies that Google will sooner or later make available.

NOTE: this is a very interesting debate and for those who recently joined it, please, read all of it, not like me initially reading only the few comments selected by Research gate.

Skeptic.

The primary value of theories and models is to help us build betters ones. And we have a long way to go. Cohen's assumption quoted by Andrew earlier (see below) was very helpful when it was promoted by the Greeks 2,300 years ago. To attempt to map contemporary models onto these terms only hinders our ability to better understand how the brain actually does the interesting things that we are studying. (Same for the terms memory, mind, consciousness, etc. Most of the processes these terms refer to are located in the prefrontal cortex after enormous processing in posterior regions - which processes did not enter into Aristotle's models because they were not "conscious.")

****

"Categorization is the mental operation by which the brain classifies objects and events. This operation is the basis for the construction of our knowledge of the world. It is the most basic phenomenon of cognition, and consequently the most fundamental problem of cognitive science."

from the intro paper in

Cohen, H., & Lefebvre, C. (Eds.). (2005). Handbook of categorization in cognitive science. Elsevier.

I quote the above because I have yet to find a more concise, poignant, and apt expression of categorization’s fundamental importance to the entirety of the cognitive sciences - including cognitive and computational neuroscience. Unfortunately, computational neuroscience deals with categorization largely through methods like neural network models. While these are fantastically successful in many ways, they are quite fundamentally incapable (at the moment) of anything resembling an explanation for conceptual categories (rather than discrimination and other ‘learning” mechanisms). B.A. Olshausen’s contribution to the volume “20 Years of Computational Neuroscience” provides another useful, concise statement: “At the end of the day we are faced with this simple truth: No one has yet spelled out a detailed model of V1 that incorporates its true biophysical complexity and exploits this complexity to process visual information in a meaningful or useful way . The problem is not just that we lack the proper data, but that we don’t even have the right conceptual framework for thinking about what is happening.” (the entire sentence after “this simple truth” is in italics).

That volume was published in 2013, and alas qualitative advances have not occurred over the last ~year. Certainly, there are hypothetical mechanisms numerous studies have implicated, such as synchronization, but here too often there are questions that must be answered before we can even test such hypotheses: “the mechanism for the emergence of correlation, synchronization, or even nearly zero-lag synchronization (ZLS) among two or more cortical areas which do not share the same input is one of the main enigmas in neuroscience. It has been argued that nonlocal synchronization is a marker of binding activities in different cortical areas into one perceptual entity”. The second sentence is a hypothesized mechanism for perceptual categorization, but as the first points out it the mechanism itself is a puzzle.

@Robert Turner,

The 2012 reference you gave of the UC Berkeley group's research on how the distribution of categories is continuous across the brain's cortex is very illuminating! I am enjoying the paper. I can appreciate it, having done a lot of mathematical data analysis (in other applications). I highly recommend it. The phenomenon is disarmingly elegant. The mathematics can, perhaps, be challenging to people without the requisite mathematical background. I still need to understand how they did linear regression of the BOLD (blood-oxygen-level-dependent) of a fMRI voxel over the 1,700+ categories. Do I have it right? I will see if I can get a more complete description of this part of the process. The other aspects of their paper are quite understandable.

I expect this will be a landmark paper.

Ten years ago processing some recordings from striatum we found that action potentials were not stereotyped events. In addition their spatial propagation is correlated with behavioral semantics

The paper was rejected almost everywhere, it was an unexpected event, against the mainstream rule of temporal patterns analysis

Reading the Neural Code: What do Spikes Mean for Behavior? Available from Nature Precedings

Dorian: "Latter using intracranial recordings from medial temporal lobe (MTL) of epileptic patients and applying the same technique I found that such transient charge density dynamics [spike directivity?] within action potentials provides better results in discriminating different categories of visual object recognition."

If spike directivity is an essential brain event underlying object recognition, how does the brain itself compute the directivity of its patterns of transient charges?

John: "Categorization is the mental operation by which the brain classifies objects and events. This operation is the basis for the construction of our knowledge of the world. It is the most basic phenomenon of cognition, and consequently the most fundamental problem of cognitive science."

If we consider the problem of knowing about the world, I think *the most fundamental problem of cognitive science* is "How do we have an experience of the space we live in, the world around us?" -- the problem of subjectivity. Classifying objects and events in the world depends first on our experiencing a world in which objects and events can exist. For a discussion of this problem, see "Where Am I ? Redux" on my RG page. For an example of categorization by a proposed system of neuronal mechanisms, see "Sparse Coding of Faces in a Neuronal Model: ..." on RG.

That's an excellent question.

Every experiment provides only a fragmented image of reality. The directivity [ spike directivity] is just a "tool" to understand what happens during an action potential, nothing more. If one can compute [ spike directivity] it doesn't mean that [ spike directivity] is "computed" by the brain.

Indeed, the brain functions as a whole, no separatness, that's the main reason we tried to develop a different theoretical model of computation ( interacting particles, charges - non-Turing , neuroelectrodynamics, NED). Since particles can be seen as "string vibrations" in a higher dimensional space, they make the harmony of cosmos and the "music of life"

NED is one possible explanation for experimental observation [ spike directivity] , however I'm open to accept any other plausible hypothesis.

First of all, i would like to thank all the contributors, whose comments on this intriguing question are both fundamental and inspiring. To come back to the neurological basis of category formation, it is important to consider a theory which does not only offer an explanation but also takes the structural order of the human brain into account. A very convincing, but not even new theory is the one presented by Prof. Larkum in the following paper:

A cellular mechanism for cortical associations: an organizing principle for the cerebral cortex. Trends Neurosci. 2013 Mar;36(3):141-51. doi: 10.1016/j.tins.2012.11.006. Epub 2012 Dec 25.

Here, it is pointed out that the 6-layered structure of the cortex might allow massive parallel computing of sensory information, based on its main cell type, the pyramidal cell. In this theory, the pyramidal cell is understood as an associative element, computing feed-forward and feedback input at the cellular level. This way, different perceptions trigger different cells with a different strength, allowing the mind to apply categories. For example, in a jungle environment, someone might be most concerned about the recognition of the category "tiger" or "predator". Cells responding to corresponding stimuli are easier stimulated due to a pre-depolarization by higher ranking computing cells. I like this model a lot due to its simplicity.

Kind regards,

Lukas

Lukas,

The argument in the Larkum paper is consistent with the synaptic matrix model simulated in "Self-Directed Learning in a Complex Environment" (available on my RG page). The most likely candidates for filter cells in the synaptic matrix are pyramidal cells.

Dear Dr. Trehub

Regarding your comment "I think *the most fundamental problem of cognitive science* is 'How do we have an experience of the space we live in, the world around us?'"- I largely agree. However, I think that the methods by which and through which we categorize our sensory-perceptual experiences is, if not equivalent to, than certainly fundamentally integral to categorization. The world around us is interpreted via our categorization and classification of stimuli into meaningful units. We see books instead of pages bound together or some conglomeration of wood, paper, and cardboard that makes up a bookshelf because of the mechanisms through which we classify books as "wholes" into one category that is neither equivalent to the papers, ink, and binding it is made up of nor the combination of this with the wooden bookshelf the books are found upon/within. Our subjective experience is the categorization of sensory experiences into, according to, and influencing our internal networks of conceptual representations. Visual experience is relatively easy to model in particular ways: we have the electromagnetic spectrum, rods and cones, the occipital lobe and primary visual pathways among many other relatively simplistic models. I say relatively because modelling neuronal reactions to visual stimuli incredibly simple compared to the ways in which our conceptual framework and cognitive processors classify, cluster, and categorize what is no more than spike trains caused by electromagnetic interactions into conceptually rich and ordered experiences. Categorization is what makes our visual/perceptual experiences different from how images are "understood" by a cell phone digital camera. Sensory experiences are a mess of signals and categorization is our experience of these in the ordered ways we interpret our experiences.

Andrew: "Categorization is what makes our visual/perceptual experiences different from how images are "understood" by a cell phone digital camera."

I agree that categorization is significantly different from the image representations in a digital camera. Cameras do not understand anything. But categorization, as such, is just one part of our perceptual experience. In fact we are unable to perceive, categorize, and understand any *aspect of the world* without the brain mechanisms of subjectivity, the egocentric representation of what is around us. When we experience the meaning of a categorized object or event it is a phenomenal experience of something in the context of a *situation* -- an experience of what is categorized in terms of salient world factors. For example, if we categorize a stimulus as a red light, its meaning varies decisively if it is a traffic signal while we are driving, as contrasted with the same signal seen from a bench in a nearby park. This is why we need the neuronal mechanisms of the retinoid system in order to understand the world.

Something that may help shed light on this topic is to consider the opposite of categorization, as a neurological phenomenon.

To me this opposite phenomenon (or phenomena) occurs in various arts and in various forms of "mystical" experiences.

Under these types of neurological states, the artist or the mystic see the stimuli as having texture, form, hue, animation ..., that is overlooked (and often invisible) to people who are busy categorizing what they see to be used/not used for some utilitarian purpose.

Often, it can be the same person, but this person is able to perceive/connect with the world differently at different times, according to his/her current situation or disposition.

Most papers in neuroscience ( e.g. the Larkum's paper) do not propose new theoretical models, they just reinforce previous theories.

Lukas is right the paper is excellent it reveals all required ingredients from all three models:

(i) Originally shown by Lord Adrian firing rate changes and other temporal patterns (ISI, STDP ) can provide information regarding categories

(ii) Synaptic activities are formally represented by connectionist (Hebbian) models, they provide information regarding categories . However, a similar connectionist model can be implemented for the Solar system, some “weights” can represent gravitational interactions within the system, providing a "dynamic balance between excitation and inhibition". In this case the resulting “synaptic matrix” tell us less than the Kepler’s laws of planetary motion.

If such “synaptic matrix” would have been the answer for category formation then we would have modeled the brain on a digital computer at least three decade ago.

(iii) In addition to synapses the entire neuron is involved (dendrites, axonal branches, ion fluxes) and the local field potential carry information regarding various categories (neuroelectrodynamics -NED). A change in spike directivity during an action potential reflects the “activation” of different synapses, it is influenced by local field potential (LFP) and also changes the electrical field.

From a computational point of view NED has more power than the first two models and can provide the answer you seek regarding spatial distribution of memory fragments within neurons, synaptic activities, category formation http://www.ncbi.nlm.nih.gov/pubmed/22480985 and semantics http://dx.doi.org/10.1038/npre.2007.61.1

Unfortunately we do not have yet the technology to compute using particle (charges). We can only simulate such processes using digital computers which is far less. However, this may change fast one day

Last week I had to drive to San Diego - a very long trip. I could not avoid and hit a piece of steel kicked up by a truck in front of me, and remained stuck on the side of the road for few hours while the “mainstream” traffic was moving along. No one paid attention except a young fellow who stopped by and offered his help to change the flat tire. His name is Nathaniel Sloan and to me he is a hero.

I’m confident that we will have soon the technology to compute using particles (charges). Someone above the “mainstream” crowd will realize the opportunity and everything in computational science may change in a blink of an eye.

In my opinion it is very difficult to raise the question of the neurological basis for category formation if we don't take into account the general rules of the brain functioning and all the data supporting the 'embodied cognition theories'. These data have clearly shown : (a) that a continuum exists between perceptual-motor activities and cortical representations; (b) that different sources of knowledge have a different weight in the representation of different categories of knowledge and (c) that the brain structures involved in the representation of dfifferent conceptual categories are located very near to the areas of convergence of these 'sources of knowledge'.

Obviously, this does not means that embodied brain mechanisms subsume all kinds and all steps of category formation. These mechanisms are more important in the formation of concrete than of abstract categories and in laying the foundations of these categories than in achieving their most advanced organization. Developmental studies have shown that in the early age the increase of knowledge and of the categorization process, depend on perceptual processing and active manipulation, and that the child acquires his first concepts observing the perceptual attributes and performing various kinds of actions with objects. At a later stage the verbal labeling and the construction of complex verbal sentences allows the abstract organization of concepts in more complex structures. Furthermore, parts of the representation of concrete concepts might derive from the similarity among members of the corresponding category and the dissimilarity to members of competing categories, whereas other parts of their representation might derive from language-mediated abstract information. For instance viewing moving beings whose shape comprise a head, a trunk and four legs could lead to the perceptual categorization of an animal (in the more abstract acceptation of a mammal), whereas only a set of complex language-mediated abstract information could lead to include in the same category the shapes of ‘a whale’, a ‘dolphin’ or of a ‘bat’ , that, on the basis of a purely perceptual categorization should be included respectively in the categories of ‘fishes’ and ‘birds’.

To me, this is an (the?) essential question,

Can be attached to the eternal problem « same or different? » behind any contrast enhancing process. « Evident" in the functional organization of the perceptual fields in the retina, and also other sensory modalities, if I understand what Arnold is proposing. With a simple rule « if two similar units are near one another, then they can be grouped, « if not, then they are part of different entities ». Same rule if they are in line, as in Kanizsa triangles or illusory contours.

This principle is also considered active in the social domaine as « métacontraste », a PhD thesis achieved by Laurent Salzarulo in 2006 (http://doc.rero.ch/record/6172?ln=fr).

The balance between «same » vs « different » can be observed in many historical steps such as the continuity-discontinuity between two nerve cells, etc.

... then the question is as how to relate these two discrete objects, and the story of the synapse proposes that an anatomical border is counterbalanced by a process or function.

To raise the « spatial dimension », one can insist on the fact that episodic memory, from its spatial memory origin by the hippocampus, stores the unique properties of the subject’s situation in space and time (the what where when questions of memory). This makes an original episode, not to be confounded with any other even though similar ones (time dimension), but of course also processed along a semantic axis. At this point, things become highly speculative!

In sum, the common property of this question, from the cellular to social level, points a very fundamental law... one that does not require an observer to remain effective... speculation again!

In this question we can find the key for understanding qualia. If an instance of a category is established, an object emerges in the mind.

I think there are two possibilities: First we realize the object and similar objects are forming a category or a new object creates a category and time after time instances are summarized under this category. If some slight differences are realized the category will split. For small children, all animals with four legs are dogs. A bit later cats are discriminated from dogs and other animals too.

Wilfried: "If an instance of a category is established, an object emerges in the mind."

Categories are actually constructed by the cognitive brain when we find it adaptive to parse our global phenomenal experience into particular sub-patterns/objects that are indexed by their coupled class-cells in synaptic matrices. Discharge of a class-cell evokes the image of its associated object in the mosaic array. When the image is back-projected into retinoid space it is consciously experienced ("emerges in the mind") as a quale, -- *something somewhere in relation to oneself* (I!). See "Learning, Imagery, Tokens and Types: The Synaptic Matrix", "Space, self, and the theater of consciousness", and "Self-Directed Learning in a Complex Environment" on RG.

Arnold,

your model is great. I agree!

Consider that categorization is the result of rule-establishing processes involving the cerebellum and prefrontal cortex (see J. Balsters et al's recent brain-imaging on this).

Such "rule-establishing processes" have been proposed for over 60 years. They have completely failed to produce anything that is comparable to "simple' mammalian brains (or brains in general), the general approach is based upon a theoretic position that predates most of our empirical study of neural dynamics and neuronal networks, and there exists no scientific model that can bridge the gap between computational neuroscience and cognitive neuroscience (let alone cognitive psychology in general).. However, thanks largely to ideological and historical reasons (the emergence of the cognitive sciences from combinatorial generative linguistics and the burgeoning computational sciences), we retain an algorithmic approach to conceptual categorization while simultaneously using models that are wholly incapable of any such processes and are based upon the procedural/syntactical manipulations of the type sea slugs are capable of but which are vastly out stripped by any living system with a brain.

Of course, it need not be true (although I think it to be so, and ridiculously obviously so) that such procedural learning is qualitatively different from categorization, but "[r]ecalling that modern digital computers...are not able to reproduce [the brain's] main functions, we are lead to conclude that the brain should work in a way fundamentally different from digital information processing."

Manrubia, S. C., Mikhailov, A. S., & Zannette, D. H. (2004). Emergence of dynamical order: synchronization phenomena in complex systems. World Scientific.

The enormous strides we've made in computational intelligence are amazing, but they are not amazing as models of living systems or of brains (let along the "mind"). "Rule-based" approaches (i.e., algorithmic) are rooted in an understanding based upon a view of physics and mathematics that Frege, Boole, Peano, Russell Whitehead, Hilbert, and Turing all thought THE necessary step in mathematics. However, the entire goal was proved impossible while Lorentz and others showed how complex "simple" systems were. They are not algorithmic, and there exist proofs (albeit disputed) that living systems cannot have computable models. Even if the brain is an algorithmic, Turing-equivalent processor, we are nowhere near to any model of categorization based upon models of cortical functions.

Current ideas in computational neuroscience regarding (ISI, STDP http://en.wikipedia.org/wiki/Spike-timing-dependent_plasticity) are just extensions of Adrian’s model (firing rate, 1914 http://en.wikipedia.org/wiki/Edgar_Adrian,_1st_Baron_Adrian).

At this point, you'll understand that the mainstream in computational neuroscience preserves a dogma older than the Turing's paradigm (1936) . Taken from this perspective one may explain why the neurological basis for category formation and many other processes are missing and possibly inaccurately described.

I didn't know they were doing brain-imaging 60 years ago. Look into it.

60 years ago they were using already using hemodynamics for whole brain functional analyses using radiographic imaging.

@Larry Vandervert

You seem to be in a much better position than I to appreciate whatever reasons on such as yourself might have for possessing an expertise that is academic in orientation and yet is underrepresented in academic literature. It seems we have both produced more work in the private sector and/or via consulting than in academia (although your contributions to both far exceed my own). Also, I have a more than strictly personal experience here: my grandfather was one of the first wave of Conant fellows, received his Ph.D. at age 25 (and wrote his dissertation on laryngeal theory in Latin) debated with Chomsky when the cognitive sciences were in their infancy, and edited the standard English reference grammar of classical Greek, yet produced most of whatever scholarship he did as a hobby while working for US embassies in Austria, Iceland, etc. I wasted several years thanks to my arrogant assumption that I, as one freshly finished with coursework, could finish a doctoral dissertation showing that quantum consciousness/quantum theories of mind were impossible despite the limits of our knowledge of the brain and the limits of our ability to know the limits of our understanding of quantum systems. For me research consulting was necessary because of blind ambition, while you have contributed to academia despite (it seems) focusing on work in the private sector.

Ironically, it is both for your accomplishments and your expertise that I find your position doubly curious. On the one hand, there’s your insultingly dismissive hortative aside that I "look into" your mischaracterization you could have corrected by glancing through elementary reference material and that, it seems, you proffered because you think it relevant that neuroimaging today somehow makes the slightest difference here (despite ~30 years of hundreds upon hundreds of neuroimaging studies that have supported incompatible theories regarding the foundations of cognition). You insult my suggestion that the explanation you propose is over half a century old and produced so little, yet I have worked almost exclusively with those who were among the founders of cognitive science and who were instrumental in developing the view you appear to find most supported. On the other hand, despite a clearly more extensive career and far more research experience than I, you too have nonetheless contributed relatively little to the cognitive sciences and yet your company sports a name suggestive of an approach virtually absent in neuroscience (thanks in no small part to neuroimaging, which deals with nonlinear systems and high dimensional data but is largely carried out by those who rely on software to run analyses they cannot understand because the curricula for psychology, social psychology, and related disciplines still does not require a sufficient background in mathematics; instead, the push-button approach dominates). However, given my lack of anything remotely related to some acknowledge expertise, perhaps I should follow your lead an offer sources, so that you can tell those people who have dedicated more years than I to "look into" their own fields. Fortunately for me, I’ve already done this to some extent here. But rather than simply regurgitate it I'll supply a few quotes and more appropriate relevant sources first (as, among other things, my previous contribution was not specific to this topic but was more general). The first is more to explain my curiosity at your position than it is meant to be informative as it is at a level far too elementary for you to gain anything from it. I have deliberately excluded all citations of basic neuroimgaging texts for beginning grad students or advanced undergrads so that you may have the opportunity to "look into it" before I leave mere literature behind and discuss the actual use of neuroimaging technologies from my own experiences using them.

"Complexity is a scientific theory that asserts that some systems display behavioural phenomena completely inexplicable by any conventional analysis of the systems’ constituent parts.

Besides, emergence refers to the appearance of higher-level properties and behaviours of a system that while obviously originating from the collective dynamics of that system’s components -are neither to be found in nor are directly deductable from the lower-level properties of that system. Emergent properties are properties of the ‘whole’ that are not possessed by any of the individual parts making up that whole...Moreover, it is becoming a commonplace that, if the 20th was the century of physics, the 21st will be the century of biology, and, more specifically, mathematical biology"

“no formal system is able to generate anything even remotely mind-like. The asymmetry between the brain and the computer is complete, all comparisons are flawed, and the idea of a computer-generated consciousness is nonsense.”

Torey, Z. (2009). The crucible of consciousness: An integrated theory of mind and brain. MIT press.

"Why would the mind work like a computer? This book is aimed—like some other recent books (e.g., Kelso, 1995; Port & van Gelder, 1995; see also Fodor, 2000)—at responding to that question with the following answer: 'It doesn’t.'"

Spivey, M. (2007). The continuity of mind. Oxford University Press.

Louie, A. H. (2007). A living system must have noncomputable models. Artificial life, 13(3), 293-297.

Fiedler, K. (2011). Voodoo correlations are everywhere—not only in neuroscience. Perspectives on Psychological Science, 6(2), 163-171.

Vul, E., & Pashler, H. (2012). Voodoo and circularity errors. Neuroimage, 62(2), 945-948.

"It occurred to me that some 2-3 years of research might be redeemed somewhat from utter failure by providing sources for those interested in the literature on quantum theories of consciousness (and similar issues, such as the relevancy of QM to living systems in general or quantum-like theories of consciousness).

All but the final categories are books/volumes, though they range (and for the most part are arranged) in order from borderline-pseudoscientific sensationalist accounts to technical monographs. The final category consists of journal articles (as well as a journal largely devoted to this question). Apart from the main categories there is very little ordering, as I made the list by copying my sources (which, apart from the journal articles, are all hard-copies) and if I took the time to arrange it better I'd end up spending hours deciding what to include rather than first-found/first-chosen.

1) Sensationalist books:

Brown, J. (2000). Minds, Machines and the Multiverse: The Quest for the quantum computer. Simon and Schuster.

Wolinsky, S. (1993). Quantum Consciousness: The Guide to Experiencing Quantum Psychology. Bramble books.

2) Non-technical books

Stapp, H. P. (2009). Mind, Matter and Quantum Mechanics (3rd Ed.). Springer.

Stapp, H. P. (2011). Mindful Universe: Quantum Mechanics and the Participating Observer (2nd Ed.). Springer

Nadeau, R. (1999). The Non-Local Universe: The New Physics and Matters of the Mind. Oxford University Press.

Suarez, A., & Adams, P. (Eds.). (2012). Is Science Compatible with Free Will?: Exploring Free Will and Consciousness in the Light of Quantum Physics and Neuroscience. Springer.

Abbott, D., Davies, P. C., & Pati, A. K. (Eds.). (2008). Quantum aspects of life. World Scientific.

Pylkkänen, P. T. (2006). Mind, matter and the implicate order. Springer.

3) Volumes from Advances in Consciousness Research

Van Loocke, P. (Ed.). (2001). The physical nature of consciousness (Vol. 29). John Benjamins Publishing.

Gordon G. Globus. (2003). Quantum closures and disclosures: Thinking-together postphenomenology and quantum brain dynamics (Vol. 50). John Benjamins.

Gordon G. Globus. (2009). The transparent becoming of world: a crossing between process philosophy and quantum neurophilosophy (Vol. 77). John Benjamins.

3) Emergence: What it lacks in a physics even physicists can't call physical it makes up for as a buzzword:

Macdonald, G., & Macdonald, C. (Eds.). (2010). Emergence in Mind (Mind Association Occasional Series). Oxford University Press.

Seager, W. (2012). Natural Fabrications: Science, Emergence and Consciousness. Springer.

Murphy, N., Ellis, G. F., & O'Connor, T. (Eds.). (2009). Downward Causation and the Neurobiology of Free Will. Springer.

Clayton, P. (2004). Mind and emergence: From quantum to consciousness. Oxford.

Koons, R. C., & Bealer, G. (Eds.). (2010). The waning of materialism. Oxford University Press.

Horst, S. W. (2007). Beyond reduction: philosophy of mind and post-reductionist philosophy of science. Oxford: Oxford University Press.

4) From pretty basic to more technical:

Tuszynski, J. A. (2006). The Emerging Physics of Consciousness. Springer.

Torey, Z. (2009). The crucible of consciousness: An integrated theory of mind and brain. MIT press.

Barrett, J. A. (1999). The quantum mechanics of minds and worlds. Oxford University Press.

Green, H. S. (2000). Information theory and quantum physics: physical foundations for understanding the conscious process. Springer.

Ivancevic, V. G., & Ivancevic, T. T. (2008). Quantum leap: from Dirac and Feynman, across the Universe, to human body and mind. World Scientific.

Ivancevic, V. G., & Ivancevic, T. T. (2010). Quantum neural computation (Vol. 40). Springer.

Matta, C. F. (Ed.). Quantum Biochemistry: Electronic Structure and Biological Activity. Wiley

5) Some books on the interpretation(s) of QM, theoretical physics, and cosmology

Schlosshauer, M. A. (2007). Decoherence: and the quantum-to-classical transition. Springer.

Jaeger, G. (2009). Entanglement, information, and the interpretation of quantum mechanics. Springer.

MacKinnon, E. M. (2011). Interpreting physics: Language and the classical/quantum divide (Vol. 289). Springer.

Maudlin, T. (2011). Quantum non-locality and relativity: Metaphysical intimations of modern physics (3rd Ed.). Wiley.

Saunders, S., Barrett, J., Kent, A., & Wallace, D. (Eds.). (2010). Many Worlds?: Everett, Quantum Theory, & Reality. Oxford University Press.

Hemmick, D. L., & Shakur, A. M. (2011). Bell's Theorem and Quantum Realism: Reassessment in Light of the Schrödinger Paradox. Springer.

And finally, actual research (mostly)- journal articles

First, there's an entire journal almost wholly devoted to this issue and those like it: NeuroQuantology (www.neuroquantology.com). I'm sure some of the papers are excellent, but in my humble opinion, those who would think twice about any journals on parapsychology and/or alternative medicines which defy known physics may find this journal analogous.

Moving on-

Segalowitz, S. J. (2009). A quantum physics account of consciousness: Much less than meets the eye. Brain and cognition, 71(2), 53.

Asano, M., Ohya, M., Tanaka, Y., Basieva, I., & Khrennikov, A. (2011). Quantum-like model of brain's functioning: Decision making from decoherence. Journal of theoretical biology, 281(1), 56-64.

Khrennikov, A. (2011). Quantum-like model of processing of information in the brain based on classical electromagnetic field. Biosystems, 105(3), 250-262.

Acacio de Barros, J., & Suppes, P. (2009). Quantum mechanics, interference, and the brain. Journal of Mathematical Psychology, 53(5), 306-313.

Persinger, M. A., & Koren, S. A. (2007). A theory of neurophysics and quantum neuroscience: implications for brain function and the limits of consciousness. International Journal of Neuroscience, 117(2), 157-175.

Baars, B. J., & Edelman, D. B. (2012). Consciousness, biology and quantum hypotheses. Physics of life reviews, 9(3), 285-294.

Kurita, Y. (2005). Indispensable role of quantum theory in the brain dynamics. BioSystems, 80(3), 263-272.

Tegmark, M. (2000). Importance of quantum decoherence in brain processes. Physical Review E, 61(4), 4194.

Rosa, L. P., & Faber, J. (2004). Quantum models of the mind: Are they compatible with environment decoherence?. Physical Review E, 70(3)

Mavromatos, N. E. (2011, July). Quantum mechanical aspects of cell microtubules: science fiction or realistic possibility?. In Journal of Physics: Conference Series (Vol. 306, No. 1)

Hu, Huping, and Maoxin Wu. "Current Landscape and Future Direction of Theoretical & Experimental Quantum Brain/Mind/Consciousness Research." Journal of Consciousness Exploration & Research 1.8 (2010).

Louie, A. H. (2005). Any material realization of the (M, R)-systems must have noncomputable models. Journal of integrative neuroscience, 4(04), 423-436.

Longo, G. (2012). Incomputability in Physics and Biology†. Mathematical Structures in Computer Science, 22(5), 880-900.

Luz Cárdenas, M., Letelier, J. C., Gutierrez, C., Cornish-Bowden, A., & Soto-Andrade, J. (2010). Closure to efficient causation, computability and artificial life. Journal of theoretical biology, 263(1), 79-92.

Thaheld, F. (2003). Biological nonlocality and the mind–brain interaction problem: comments on a new empirical approach. BioSystems, 70(1), 35-41.

John, E. R. (2002). The neurophysics of consciousness. Brain Research Reviews, 39(1), 1-28.

Louie, A. H. (2007). A living system must have noncomputable models. Artificial life, 13(3), 293-297.

Smith, C. U. (2009). The ‘hard problem’and the quantum physicists. Part 2: Modern times. Brain and cognition, 71(2), 54-63.

Persinger, M. A., & Koren, S. A. (2007). A theory of neurophysics and quantum neuroscience: implications for brain function and the limits of consciousness. International Journal of Neuroscience, 117(2), 157-175.

And that's all the typing I can manage in one night. It's not exactly exhaustive, but then this isn't exactly a forum for exhaustive reviews. A final note:

For those with a fair amount of familiarity with mathematics (in particular, abstract algebras & measure theory or similar topics), I recommend Hall's Quantum Theory for Mathematicians (Graduate Texts in Mathematics). Despite the fact that we find "observables" in quantum mechanics to correspond not to what is observed but to a mathematical function, physicists like Dirac still managed to render alien the mathematical structure of QM: "Mathematicians tend to despise Dirac notation, because it can prevent them from making important distinctions, but physicists love it, because they are always forgetting that such distinctions exist and the notation liberates them from having to remember" (http://people.cs.clemson.edu/~steve/CW/395/CS483-part1.pdf)"

Hello everyone

It is a topic very interesting and related to something that I am researching now.

A simple way to categorize from the simple to the complex - at least in the binocular visual system - is by capturing stimuli ranked according to their spatial frequency in the striate cortex.

So far, this model is uni-directional, linear, still has not been uploaded to the network.

Although the input is different between information from one and other eye, the categorization follows a very simple law: the law of accion-inhibicion.

Curiously is not the input who sets the cortical hierarchy, but the ability or not to process the information, and this is something inherent in the striate cortex ocular dominance columns.

These columns have an ancestral system: prioritize information based on the spatial frequency of the stimulus.

From this raw elemental but ranked, will the action potential which determine there via the information followed. If the stimulus is able to activate receptors for color, shape, movement, contrast, etc.

The magic of the cortical network is not to do great deeds, but to know order information, prioritize it, incorporate it, and learn from this experience.

When I first entered into this discussion, I thought that it was a place where an interaction was requested between cognitivists and neuroscientists, whereas now I think that the debate is much more within cognitivists than between cognitivists and neuroscientists.

Now, a problem when cognitivists ask what is a neurological basis for category formation is that they tend to identify these neurological bases with the brain and in particular with the micro-architecture of the brain, namely with the synaptic level, somehow disregarding the fact that the brain is tightly linked to the body and that the early categories are formed in the interaction between the body (its actions and perceptual modalities) and the external world. For this reason, in my opinion, the neurological bases for category formation,

certainly require the facilitation and inhibition of the synapses belonging to sensory-motor circuits that are systematically, simultaneously activated by certain kinds of external stimuli, but mainly concern the macro-architecture of the brain, namely the convergence zones where the sources of knowledge playing a critical role in the formation of different categories anatomically converge.

The macro-architecture of the brain can provide clues about category formation, but to understand the neurological basis for category formation we need to explicate the structure and dynamics of neuronal *mechanisms*, and how they are organized into systems that can be demonstrated to have the competence to do the job. I don't see how it is possible to do this without offering models of the essential micro-architecture of the brain.

In general, any functional neuroscience research must largely abandon any computational neuroscience or any models that are otherwise at the level of synapses. Issues as fundamental as the "neural code" still remain unresolved, and thus functional neural processes such as categorization (which is perhaps THE functional neuronal process) cannot be answered using any measure of neurophysiological dynamics unless it is theory-laden and structural. In fact, you seem to point to a rather central divide within the cognitive sciences (including cognitive neurosciences) with your remark about the "interaction between the body...and the external world", which could be interpreted as supporting "embodied cognition". Chomsky, Pinker, Caramazza, etc., are all thoroughly dismissive of this theory. Choi, Lakoff, Gibbs, etc., believe it to be as "proven" as it gets in terms of science. Cognitive neuroscience is almost entirely concerned with macro-architecture (as you put it). We lack the requisite knowledge for any biologically rooted model of categorization that isn't either highly theoretical or which eschews computational neuroscience almost in its entirety. But we are closing in on the micro-level dynamics in such a way that we might at least rule out some explanatory frameworks. And I, for one, look forward to such work, and hope that I can steal it and pawn it of as my own.

From my point of view that is, the binocular visual system, this prioritizes the information in a really elegant way. The primary visual cortex ordered visual information according to the characteristics of the stimulus in a relatively short time (first level), but requires a longer time to define what you see and where you can see.

For this reason, the information of the striate cortex is sent into working memory areas adjacent to the temporal lobes (second level, - to put it in some way-). All of the above is the basic activity of the visual system. This activity requires two-way communication competent (third level): one of them the occipito-temporal for identification of Visual elements, and the other the occipito-parietal pathway, for the location of the same.

Know the capacity and efficiency of pathways and neuronal circuits, continued as a challenge for all of us, but somehow, allows us to understand how it is that the brain "prioritizes" systematically information.

Prioritize requires a substrate of anatomo-functional defined and inherent to each individual, which can be remodeled and therefore, able to improve certain visual skills.(fourth level)

The world is measured through the binocular visual system, models of objects, surfaces, three-dimensional (3D) scenes and events are constructed by vision involving perception and not only perfect optics of the image, for example.

Perception (fifth level) carries organicity and functionality, but while visual acuity depends basically of anatomo-functional integrity of the elements that make up the visual pathway, visual learning (7th level) requires the narrow link between the spatial vision and attention (6th level).

The organicity and neuro-functionality of the human binocular visual system although they are inseparable, unfortunately often analysed separately, both the Visual processes such as the structure and even the internal dialogue of the visual system (8th level). But, an issue is to understand the State neurofunctional, electric and chemical that carries the visual process and is another, understand the construct of perceptual learning through visual experience (9th level), as well as identify and even improve visual skills, and all analyzed by means of the "visual perceptual" analysis.

None of this is feasible to identify using conventional studies of neuroimaging (RMI, Rx, EEG, TAC, etc), but knowing, on the one hand, the microarchitecture by Voxel analysis of images (these studies have a high spatial resolution), and through studies neuro-functional equally fine but that they possess a high temporal resolution, for example by the neurometric analysis. Of course there are "hybrid" methods as the RMI functional, but it is the analysis of the information that allows us to correlate structure and function and conduct a good way to deal with the order in which the information is prioritized. It seems that the nervous system categorizes based on the priority in which all these elements are processed.

Unfortunately in many cases we see “what we expect to see”.

If one studies the connectivity (a broad metaphor) sooner or later he/she will discover a “connectivity” pattern for a model of categorization.

A physical model is needed to avoid such bias even for computational neuroscience.

Dorian,

if you see, what you are expecting, our mind has a problem. The association process does not fit exactly and this is an interesting situation we can learn much about the functionality of the visual cortex and its relation to the recognition process.

Human observations can be biased. Previous hypotheses and observer's beliefs are more likely to be confirmed by experimental data. Max Planck had a deep understanding of this issue. “We ourselves are part of nature and therefore part of the mystery that we are trying to solve”.

Dorian, from *Edge*:

.............................................................

Arnold Trehub

Modern science is a product of biology

The entire conceptual edifice of modern science is a product of biology. Even the most basic and profound ideas of science — think relativity, quantum theory, the theory of evolution — are generated and necessarily limited by the particular capacities of our human biology. This implies that the content and scope of scientific knowledge is not open-ended.

..............................................................

We always see, what we are expecting. If there is something new like a riddle pic, we have to wait some seconds if we are able to see clear.

Apropos this discussion, "Dissipation of dark energy by cortex in knowledge retrieval" downloadable at Walter J. Freeman's RG profile, https://www.researchgate.net/profile/Walter_Freeman2.